Add files using upload-large-folder tool

.gradio/certificate.pem

@@ -0,0 +1,31 @@
+-----BEGIN CERTIFICATE-----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=
+-----END CERTIFICATE-----

cosyvoice/utils/common.py

@@ -0,0 +1,166 @@
+# Copyright (c) 2020 Mobvoi Inc (Binbin Zhang)
+#               2024 Alibaba Inc (authors: Xiang Lyu)
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#   http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# Modified from ESPnet(https://github.com/espnet/espnet)
+"""Unility functions for Transformer."""
+
+import random
+from typing import List
+
+import numpy as np
+import torch
+
+IGNORE_ID = -1
+
+
+def pad_list(xs: List[torch.Tensor], pad_value: int):
+    """Perform padding for the list of tensors.
+
+    Args:
+        xs (List): List of Tensors [(T_1, `*`), (T_2, `*`), ..., (T_B, `*`)].
+        pad_value (float): Value for padding.
+
+    Returns:
+        Tensor: Padded tensor (B, Tmax, `*`).
+
+    Examples:
+        >>> x = [torch.ones(4), torch.ones(2), torch.ones(1)]
+        >>> x
+        [tensor([1., 1., 1., 1.]), tensor([1., 1.]), tensor([1.])]
+        >>> pad_list(x, 0)
+        tensor([[1., 1., 1., 1.],
+                [1., 1., 0., 0.],
+                [1., 0., 0., 0.]])
+
+    """
+    max_len = max([len(item) for item in xs])
+    batchs = len(xs)
+    ndim = xs[0].ndim
+    if ndim == 1:
+        pad_res = torch.zeros(batchs,
+                              max_len,
+                              dtype=xs[0].dtype,
+                              device=xs[0].device)
+    elif ndim == 2:
+        pad_res = torch.zeros(batchs,
+                              max_len,
+                              xs[0].shape[1],
+                              dtype=xs[0].dtype,
+                              device=xs[0].device)
+    elif ndim == 3:
+        pad_res = torch.zeros(batchs,
+                              max_len,
+                              xs[0].shape[1],
+                              xs[0].shape[2],
+                              dtype=xs[0].dtype,
+                              device=xs[0].device)
+    else:
+        raise ValueError(f"Unsupported ndim: {ndim}")
+    pad_res.fill_(pad_value)
+    for i in range(batchs):
+        pad_res[i, :len(xs[i])] = xs[i]
+    return pad_res
+
+
+def th_accuracy(pad_outputs: torch.Tensor, pad_targets: torch.Tensor,
+                ignore_label: int) -> torch.Tensor:
+    """Calculate accuracy.
+
+    Args:
+        pad_outputs (Tensor): Prediction tensors (B * Lmax, D).
+        pad_targets (LongTensor): Target label tensors (B, Lmax).
+        ignore_label (int): Ignore label id.
+
+    Returns:
+        torch.Tensor: Accuracy value (0.0 - 1.0).
+
+    """
+    pad_pred = pad_outputs.view(pad_targets.size(0), pad_targets.size(1),
+                                pad_outputs.size(1)).argmax(2)
+    mask = pad_targets != ignore_label
+    numerator = torch.sum(
+        pad_pred.masked_select(mask) == pad_targets.masked_select(mask))
+    denominator = torch.sum(mask)
+    return (numerator / denominator).detach()
+
+
+def get_padding(kernel_size, dilation=1):
+    return int((kernel_size * dilation - dilation) / 2)
+
+
+def init_weights(m, mean=0.0, std=0.01):
+    classname = m.__class__.__name__
+    if classname.find("Conv") != -1:
+        m.weight.data.normal_(mean, std)
+
+
+# Repetition Aware Sampling in VALL-E 2
+def ras_sampling(weighted_scores, decoded_tokens, sampling, top_p=0.8, top_k=25, win_size=10, tau_r=0.1):
+    top_ids = nucleus_sampling(weighted_scores, top_p=top_p, top_k=top_k)
+    rep_num = (torch.tensor(decoded_tokens[-win_size:]).to(weighted_scores.device) == top_ids).sum().item()
+    if rep_num >= win_size * tau_r:
+        top_ids = random_sampling(weighted_scores, decoded_tokens, sampling)
+    return top_ids
+
+
+def nucleus_sampling(weighted_scores, top_p=0.8, top_k=25):
+    prob, indices = [], []
+    cum_prob = 0.0
+    sorted_value, sorted_idx = weighted_scores.softmax(dim=0).sort(descending=True, stable=True)
+    for i in range(len(sorted_idx)):
+        # sampling both top-p and numbers.
+        if cum_prob < top_p and len(prob) < top_k:
+            cum_prob += sorted_value[i]
+            prob.append(sorted_value[i])
+            indices.append(sorted_idx[i])
+        else:
+            break
+    prob = torch.tensor(prob).to(weighted_scores)
+    indices = torch.tensor(indices, dtype=torch.long).to(weighted_scores.device)
+    top_ids = indices[prob.multinomial(1, replacement=True)]
+    return top_ids
+
+
+def random_sampling(weighted_scores, decoded_tokens, sampling):
+    top_ids = weighted_scores.softmax(dim=0).multinomial(1, replacement=True)
+    return top_ids
+
+
+def fade_in_out(fade_in_mel, fade_out_mel, window):
+    device = fade_in_mel.device
+    fade_in_mel, fade_out_mel = fade_in_mel.cpu(), fade_out_mel.cpu()
+    mel_overlap_len = int(window.shape[0] / 2)
+    if fade_in_mel.device == torch.device('cpu'):
+        fade_in_mel = fade_in_mel.clone()
+    fade_in_mel[..., :mel_overlap_len] = fade_in_mel[..., :mel_overlap_len] * window[:mel_overlap_len] + \
+        fade_out_mel[..., -mel_overlap_len:] * window[mel_overlap_len:]
+    return fade_in_mel.to(device)
+
+
+def set_all_random_seed(seed):
+    random.seed(seed)
+    np.random.seed(seed)
+    torch.manual_seed(seed)
+    torch.cuda.manual_seed_all(seed)
+
+
+def mask_to_bias(mask: torch.Tensor, dtype: torch.dtype) -> torch.Tensor:
+    assert mask.dtype == torch.bool
+    assert dtype in [torch.float32, torch.bfloat16, torch.float16]
+    mask = mask.to(dtype)
+    # attention mask bias
+    # NOTE(Mddct): torch.finfo jit issues
+    #     chunk_masks = (1.0 - chunk_masks) * torch.finfo(dtype).min
+    mask = (1.0 - mask) * -1.0e+10
+    return mask

cosyvoice/utils/losses.py

@@ -0,0 +1,20 @@
+import torch
+import torch.nn.functional as F
+
+
+def tpr_loss(disc_real_outputs, disc_generated_outputs, tau):
+    loss = 0
+    for dr, dg in zip(disc_real_outputs, disc_generated_outputs):
+        m_DG = torch.median((dr - dg))
+        L_rel = torch.mean((((dr - dg) - m_DG) ** 2)[dr < dg + m_DG])
+        loss += tau - F.relu(tau - L_rel)
+    return loss
+
+
+def mel_loss(real_speech, generated_speech, mel_transforms):
+    loss = 0
+    for transform in mel_transforms:
+        mel_r = transform(real_speech)
+        mel_g = transform(generated_speech)
+        loss += F.l1_loss(mel_g, mel_r)
+    return loss

examples/libritts/cosyvoice/conf/cosyvoice.fromscratch.yaml

@@ -0,0 +1,257 @@
+# set random seed, so that you may reproduce your result.
+__set_seed1: !apply:random.seed [1986]
+__set_seed2: !apply:numpy.random.seed [1986]
+__set_seed3: !apply:torch.manual_seed [1986]
+__set_seed4: !apply:torch.cuda.manual_seed_all [1986]
+
+# fixed params
+sample_rate: 22050
+text_encoder_input_size: 512
+llm_input_size: 1024
+llm_output_size: 1024
+spk_embed_dim: 192
+
+# model params
+# for all class/function included in this repo, we use !<name> or !<new> for intialization, so that user may find all corresponding class/function according to one single yaml.
+# for system/third_party class/function, we do not require this.
+llm: !new:cosyvoice.llm.llm.TransformerLM
+    text_encoder_input_size: !ref <text_encoder_input_size>
+    llm_input_size: !ref <llm_input_size>
+    llm_output_size: !ref <llm_output_size>
+    text_token_size: 51866 # change to 60515 if you want to train with CosyVoice-300M-25Hz recipe
+    speech_token_size: 4096
+    length_normalized_loss: True
+    lsm_weight: 0
+    spk_embed_dim: !ref <spk_embed_dim>
+    text_encoder: !new:cosyvoice.transformer.encoder.ConformerEncoder
+        input_size: !ref <text_encoder_input_size>
+        output_size: 1024
+        attention_heads: 8
+        linear_units: 2048
+        num_blocks: 3
+        dropout_rate: 0.1
+        positional_dropout_rate: 0.1
+        attention_dropout_rate: 0.0
+        normalize_before: True
+        input_layer: 'linear'
+        pos_enc_layer_type: 'rel_pos_espnet'
+        selfattention_layer_type: 'rel_selfattn'
+        use_cnn_module: False
+        macaron_style: False
+        use_dynamic_chunk: False
+        use_dynamic_left_chunk: False
+        static_chunk_size: 1
+    llm: !new:cosyvoice.transformer.encoder.TransformerEncoder
+        input_size: !ref <llm_input_size>
+        output_size: !ref <llm_output_size>
+        attention_heads: 8
+        linear_units: 2048
+        num_blocks: 7
+        dropout_rate: 0.1
+        positional_dropout_rate: 0.1
+        attention_dropout_rate: 0.0
+        input_layer: 'linear_legacy'
+        pos_enc_layer_type: 'rel_pos_espnet'
+        selfattention_layer_type: 'rel_selfattn'
+        static_chunk_size: 1
+    sampling: !name:cosyvoice.utils.common.ras_sampling
+        top_p: 0.8
+        top_k: 25
+        win_size: 10
+        tau_r: 0.1
+
+flow: !new:cosyvoice.flow.flow.MaskedDiffWithXvec
+    input_size: 512
+    output_size: 80
+    spk_embed_dim: !ref <spk_embed_dim>
+    output_type: 'mel'
+    vocab_size: 4096
+    input_frame_rate: 50 # change to 25 if you want to train with CosyVoice-300M-25Hz recipe
+    only_mask_loss: True
+    encoder: !new:cosyvoice.transformer.encoder.ConformerEncoder
+        output_size: 512
+        attention_heads: 4
+        linear_units: 1024
+        num_blocks: 3
+        dropout_rate: 0.1
+        positional_dropout_rate: 0.1
+        attention_dropout_rate: 0.1
+        normalize_before: True
+        input_layer: 'linear'
+        pos_enc_layer_type: 'rel_pos_espnet'
+        selfattention_layer_type: 'rel_selfattn'
+        input_size: 512
+        use_cnn_module: False
+        macaron_style: False
+    length_regulator: !new:cosyvoice.flow.length_regulator.InterpolateRegulator
+        channels: 80
+        sampling_ratios: [1, 1, 1, 1]
+    decoder: !new:cosyvoice.flow.flow_matching.ConditionalCFM
+        in_channels: 240
+        n_spks: 1
+        spk_emb_dim: 80
+        cfm_params: !new:omegaconf.DictConfig
+            content:
+                sigma_min: 1e-06
+                solver: 'euler'
+                t_scheduler: 'cosine'
+                training_cfg_rate: 0.2
+                inference_cfg_rate: 0.7
+                reg_loss_type: 'l1'
+        estimator: !new:cosyvoice.flow.decoder.ConditionalDecoder
+            in_channels: 320
+            out_channels: 80
+            channels: [256, 256]
+            dropout: 0.0
+            attention_head_dim: 64
+            n_blocks: 4
+            num_mid_blocks: 8
+            num_heads: 8
+            act_fn: 'gelu'
+
+hift: !new:cosyvoice.hifigan.generator.HiFTGenerator
+    in_channels: 80
+    base_channels: 512
+    nb_harmonics: 8
+    sampling_rate: !ref <sample_rate>
+    nsf_alpha: 0.1
+    nsf_sigma: 0.003
+    nsf_voiced_threshold: 10
+    upsample_rates: [8, 8]
+    upsample_kernel_sizes: [16, 16]
+    istft_params:
+        n_fft: 16
+        hop_len: 4
+    resblock_kernel_sizes: [3, 7, 11]
+    resblock_dilation_sizes: [[1, 3, 5], [1, 3, 5], [1, 3, 5]]
+    source_resblock_kernel_sizes: [7, 11]
+    source_resblock_dilation_sizes: [[1, 3, 5], [1, 3, 5]]
+    lrelu_slope: 0.1
+    audio_limit: 0.99
+    f0_predictor: !new:cosyvoice.hifigan.f0_predictor.ConvRNNF0Predictor
+        num_class: 1
+        in_channels: 80
+        cond_channels: 512
+
+# gan related module
+mel_spec_transform1: !name:matcha.utils.audio.mel_spectrogram
+    n_fft: 1024
+    num_mels: 80
+    sampling_rate: !ref <sample_rate>
+    hop_size: 256
+    win_size: 1024
+    fmin: 0
+    fmax: null
+    center: False
+hifigan: !new:cosyvoice.hifigan.hifigan.HiFiGan
+    generator: !ref <hift>
+    discriminator: !new:cosyvoice.hifigan.discriminator.MultipleDiscriminator
+        mpd: !new:matcha.hifigan.models.MultiPeriodDiscriminator
+        mrd: !new:cosyvoice.hifigan.discriminator.MultiResolutionDiscriminator
+    mel_spec_transform: [
+        !ref <mel_spec_transform1>
+    ]
+
+# processor functions
+parquet_opener: !name:cosyvoice.dataset.processor.parquet_opener
+get_tokenizer: !name:whisper.tokenizer.get_tokenizer # change to !name:cosyvoice.tokenizer.tokenizer.get_tokenizer if you want to train with CosyVoice-300M-25Hz recipe
+    multilingual: True
+    num_languages: 100
+    language: 'en'
+    task: 'transcribe'
+allowed_special: 'all'
+tokenize: !name:cosyvoice.dataset.processor.tokenize
+    get_tokenizer: !ref <get_tokenizer>
+    allowed_special: !ref <allowed_special>
+filter: !name:cosyvoice.dataset.processor.filter
+    max_length: 40960
+    min_length: 0
+    token_max_length: 200
+    token_min_length: 1
+resample: !name:cosyvoice.dataset.processor.resample
+    resample_rate: !ref <sample_rate>
+truncate: !name:cosyvoice.dataset.processor.truncate
+    truncate_length: 24576 # must be a multiplier of hop_size
+feat_extractor: !name:matcha.utils.audio.mel_spectrogram
+    n_fft: 1024
+    num_mels: 80
+    sampling_rate: !ref <sample_rate>
+    hop_size: 256
+    win_size: 1024
+    fmin: 0
+    fmax: 8000
+    center: False
+compute_fbank: !name:cosyvoice.dataset.processor.compute_fbank
+    feat_extractor: !ref <feat_extractor>
+compute_f0: !name:cosyvoice.dataset.processor.compute_f0
+    sample_rate: !ref <sample_rate>
+    hop_size: 256
+parse_embedding: !name:cosyvoice.dataset.processor.parse_embedding
+    normalize: True
+shuffle: !name:cosyvoice.dataset.processor.shuffle
+    shuffle_size: 1000
+sort: !name:cosyvoice.dataset.processor.sort
+    sort_size: 500  # sort_size should be less than shuffle_size
+batch: !name:cosyvoice.dataset.processor.batch
+    batch_type: 'dynamic'
+    max_frames_in_batch: 12000
+padding: !name:cosyvoice.dataset.processor.padding
+    use_spk_embedding: False # change to True during sft
+
+# dataset processor pipeline
+data_pipeline: [
+    !ref <parquet_opener>,
+    !ref <tokenize>,
+    !ref <filter>,
+    !ref <resample>,
+    !ref <compute_fbank>,
+    !ref <parse_embedding>,
+    !ref <shuffle>,
+    !ref <sort>,
+    !ref <batch>,
+    !ref <padding>,
+]
+data_pipeline_gan: [
+    !ref <parquet_opener>,
+    !ref <tokenize>,
+    !ref <filter>,
+    !ref <resample>,
+    !ref <truncate>,
+    !ref <compute_fbank>,
+    !ref <compute_f0>,
+    !ref <parse_embedding>,
+    !ref <shuffle>,
+    !ref <sort>,
+    !ref <batch>,
+    !ref <padding>,
+]
+
+# llm flow train conf
+train_conf:
+    optim: adam
+    optim_conf:
+        lr: 0.002 # change to 0.001 if you want to train flow from scratch
+    scheduler: warmuplr
+    scheduler_conf:
+        warmup_steps: 25000
+    max_epoch: 200
+    grad_clip: 5
+    accum_grad: 2
+    log_interval: 100
+    save_per_step: -1
+
+# gan train conf
+train_conf_gan:
+    optim: adam
+    optim_conf:
+        lr: 0.0002 # use small lr for gan training
+    scheduler: constantlr
+    optim_d: adam
+    optim_conf_d:
+        lr: 0.0002 # use small lr for gan training
+    scheduler_d: constantlr
+    max_epoch: 200
+    grad_clip: 5
+    accum_grad: 1 # in gan training, accum_grad must be 1
+    log_interval: 100
+    save_per_step: -1

examples/libritts/cosyvoice/conf/ds_stage2.json

@@ -0,0 +1,42 @@
+{
+  "train_micro_batch_size_per_gpu": 1,
+  "gradient_accumulation_steps": 1,
+  "steps_per_print": 100,
+  "gradient_clipping": 5,
+  "fp16": {
+    "enabled": false,
+    "auto_cast": false,
+    "loss_scale": 0,
+    "initial_scale_power": 16,
+    "loss_scale_window": 256,
+    "hysteresis": 2,
+    "consecutive_hysteresis": false,
+    "min_loss_scale": 1
+  },
+  "bf16": {
+    "enabled": false
+  },
+  "zero_force_ds_cpu_optimizer": false,
+  "zero_optimization": {
+    "stage": 2,
+    "offload_optimizer": {
+      "device": "none",
+      "pin_memory": true
+    },
+    "allgather_partitions": true,
+    "allgather_bucket_size": 5e8,
+    "overlap_comm": false,
+    "reduce_scatter": true,
+    "reduce_bucket_size": 5e8,
+    "contiguous_gradients" : true
+  },
+  "optimizer": {
+    "type": "AdamW",
+    "params": {
+        "lr": 0.001,
+        "weight_decay": 0.0001,
+        "torch_adam": true,
+        "adam_w_mode": true
+    }
+  }
+}

examples/libritts/cosyvoice/tts_text.json

@@ -0,0 +1,5 @@
+{
+  "1089_134686_000002_000000": [
+    "hello, my name is Jack. What is your name?"
+  ]
+}

examples/magicdata-read/cosyvoice/conf/cosyvoice.fromscratch.yaml

@@ -0,0 +1,203 @@
+# set random seed, so that you may reproduce your result.
+__set_seed1: !apply:random.seed [1986]
+__set_seed2: !apply:numpy.random.seed [1986]
+__set_seed3: !apply:torch.manual_seed [1986]
+__set_seed4: !apply:torch.cuda.manual_seed_all [1986]
+
+# fixed params
+sample_rate: 22050
+text_encoder_input_size: 512
+llm_input_size: 1024
+llm_output_size: 1024
+spk_embed_dim: 192
+
+# model params
+# for all class/function included in this repo, we use !<name> or !<new> for intialization, so that user may find all corresponding class/function according to one single yaml.
+# for system/third_party class/function, we do not require this.
+llm: !new:cosyvoice.llm.llm.TransformerLM
+    text_encoder_input_size: !ref <text_encoder_input_size>
+    llm_input_size: !ref <llm_input_size>
+    llm_output_size: !ref <llm_output_size>
+    text_token_size: 51866 # change to 60515 if you want to train with CosyVoice-300M-25Hz recipe
+    speech_token_size: 4096
+    length_normalized_loss: True
+    lsm_weight: 0
+    spk_embed_dim: !ref <spk_embed_dim>
+    text_encoder: !new:cosyvoice.transformer.encoder.ConformerEncoder
+        input_size: !ref <text_encoder_input_size>
+        output_size: 1024
+        attention_heads: 8
+        linear_units: 2048
+        num_blocks: 3
+        dropout_rate: 0.1
+        positional_dropout_rate: 0.1
+        attention_dropout_rate: 0.0
+        normalize_before: True
+        input_layer: 'linear'
+        pos_enc_layer_type: 'rel_pos_espnet'
+        selfattention_layer_type: 'rel_selfattn'
+        use_cnn_module: False
+        macaron_style: False
+        use_dynamic_chunk: False
+        use_dynamic_left_chunk: False
+        static_chunk_size: 1
+    llm: !new:cosyvoice.transformer.encoder.TransformerEncoder
+        input_size: !ref <llm_input_size>
+        output_size: !ref <llm_output_size>
+        attention_heads: 8
+        linear_units: 2048
+        num_blocks: 7
+        dropout_rate: 0.1
+        positional_dropout_rate: 0.1
+        attention_dropout_rate: 0.0
+        input_layer: 'linear_legacy'
+        pos_enc_layer_type: 'rel_pos_espnet'
+        selfattention_layer_type: 'rel_selfattn'
+        static_chunk_size: 1
+    sampling: !name:cosyvoice.utils.common.ras_sampling
+        top_p: 0.8
+        top_k: 25
+        win_size: 10
+        tau_r: 0.1
+
+flow: !new:cosyvoice.flow.flow.MaskedDiffWithXvec
+    input_size: 512
+    output_size: 80
+    spk_embed_dim: !ref <spk_embed_dim>
+    output_type: 'mel'
+    vocab_size: 4096
+    input_frame_rate: 50 # change to 25 if you want to train with CosyVoice-300M-25Hz recipe
+    only_mask_loss: True
+    encoder: !new:cosyvoice.transformer.encoder.ConformerEncoder
+        output_size: 512
+        attention_heads: 4
+        linear_units: 1024
+        num_blocks: 3
+        dropout_rate: 0.1
+        positional_dropout_rate: 0.1
+        attention_dropout_rate: 0.1
+        normalize_before: True
+        input_layer: 'linear'
+        pos_enc_layer_type: 'rel_pos_espnet'
+        selfattention_layer_type: 'rel_selfattn'
+        input_size: 512
+        use_cnn_module: False
+        macaron_style: False
+    length_regulator: !new:cosyvoice.flow.length_regulator.InterpolateRegulator
+        channels: 80
+        sampling_ratios: [1, 1, 1, 1]
+    decoder: !new:cosyvoice.flow.flow_matching.ConditionalCFM
+        in_channels: 240
+        n_spks: 1
+        spk_emb_dim: 80
+        cfm_params: !new:omegaconf.DictConfig
+            content:
+                sigma_min: 1e-06
+                solver: 'euler'
+                t_scheduler: 'cosine'
+                training_cfg_rate: 0.2
+                inference_cfg_rate: 0.7
+                reg_loss_type: 'l1'
+        estimator: !new:cosyvoice.flow.decoder.ConditionalDecoder
+            in_channels: 320
+            out_channels: 80
+            channels: [256, 256]
+            dropout: 0.0
+            attention_head_dim: 64
+            n_blocks: 4
+            num_mid_blocks: 8
+            num_heads: 8
+            act_fn: 'gelu'
+
+hift: !new:cosyvoice.hifigan.generator.HiFTGenerator
+    in_channels: 80
+    base_channels: 512
+    nb_harmonics: 8
+    sampling_rate: !ref <sample_rate>
+    nsf_alpha: 0.1
+    nsf_sigma: 0.003
+    nsf_voiced_threshold: 10
+    upsample_rates: [8, 8]
+    upsample_kernel_sizes: [16, 16]
+    istft_params:
+        n_fft: 16
+        hop_len: 4
+    resblock_kernel_sizes: [3, 7, 11]
+    resblock_dilation_sizes: [[1, 3, 5], [1, 3, 5], [1, 3, 5]]
+    source_resblock_kernel_sizes: [7, 11]
+    source_resblock_dilation_sizes: [[1, 3, 5], [1, 3, 5]]
+    lrelu_slope: 0.1
+    audio_limit: 0.99
+    f0_predictor: !new:cosyvoice.hifigan.f0_predictor.ConvRNNF0Predictor
+        num_class: 1
+        in_channels: 80
+        cond_channels: 512
+
+# processor functions
+parquet_opener: !name:cosyvoice.dataset.processor.parquet_opener
+get_tokenizer: !name:whisper.tokenizer.get_tokenizer # change to !name:cosyvoice.tokenizer.tokenizer.get_tokenizer if you want to train with CosyVoice-300M-25Hz recipe
+    multilingual: True
+    num_languages: 100
+    language: 'en'
+    task: 'transcribe'
+allowed_special: 'all'
+tokenize: !name:cosyvoice.dataset.processor.tokenize
+    get_tokenizer: !ref <get_tokenizer>
+    allowed_special: !ref <allowed_special>
+filter: !name:cosyvoice.dataset.processor.filter
+    max_length: 40960
+    min_length: 0
+    token_max_length: 200
+    token_min_length: 1
+resample: !name:cosyvoice.dataset.processor.resample
+    resample_rate: !ref <sample_rate>
+feat_extractor: !name:matcha.utils.audio.mel_spectrogram
+    n_fft: 1024
+    num_mels: 80
+    sampling_rate: !ref <sample_rate>
+    hop_size: 256
+    win_size: 1024
+    fmin: 0
+    fmax: 8000
+    center: False
+compute_fbank: !name:cosyvoice.dataset.processor.compute_fbank
+    feat_extractor: !ref <feat_extractor>
+parse_embedding: !name:cosyvoice.dataset.processor.parse_embedding
+    normalize: True
+shuffle: !name:cosyvoice.dataset.processor.shuffle
+    shuffle_size: 1000
+sort: !name:cosyvoice.dataset.processor.sort
+    sort_size: 500  # sort_size should be less than shuffle_size
+batch: !name:cosyvoice.dataset.processor.batch
+    batch_type: 'dynamic'
+    max_frames_in_batch: 12000
+padding: !name:cosyvoice.dataset.processor.padding
+    use_spk_embedding: False # change to True during sft
+
+# dataset processor pipeline
+data_pipeline: [
+    !ref <parquet_opener>,
+    !ref <tokenize>,
+    !ref <filter>,
+    !ref <resample>,
+    !ref <compute_fbank>,
+    !ref <parse_embedding>,
+    !ref <shuffle>,
+    !ref <sort>,
+    !ref <batch>,
+    !ref <padding>,
+]
+
+# train conf
+train_conf:
+    optim: adam
+    optim_conf:
+        lr: 0.002 # change to 0.001 if you want to train flow from scratch
+    scheduler: warmuplr
+    scheduler_conf:
+        warmup_steps: 25000
+    max_epoch: 200
+    grad_clip: 5
+    accum_grad: 2
+    log_interval: 100
+    save_per_step: -1

examples/magicdata-read/cosyvoice/conf/cosyvoice.yaml

@@ -0,0 +1,203 @@
+# set random seed, so that you may reproduce your result.
+__set_seed1: !apply:random.seed [1986]
+__set_seed2: !apply:numpy.random.seed [1986]
+__set_seed3: !apply:torch.manual_seed [1986]
+__set_seed4: !apply:torch.cuda.manual_seed_all [1986]
+
+# fixed params
+sample_rate: 22050
+text_encoder_input_size: 512
+llm_input_size: 1024
+llm_output_size: 1024
+spk_embed_dim: 192
+
+# model params
+# for all class/function included in this repo, we use !<name> or !<new> for intialization, so that user may find all corresponding class/function according to one single yaml.
+# for system/third_party class/function, we do not require this.
+llm: !new:cosyvoice.llm.llm.TransformerLM
+    text_encoder_input_size: !ref <text_encoder_input_size>
+    llm_input_size: !ref <llm_input_size>
+    llm_output_size: !ref <llm_output_size>
+    text_token_size: 51866 # change to 60515 if you want to train with CosyVoice-300M-25Hz recipe
+    speech_token_size: 4096
+    length_normalized_loss: True
+    lsm_weight: 0
+    spk_embed_dim: !ref <spk_embed_dim>
+    text_encoder: !new:cosyvoice.transformer.encoder.ConformerEncoder
+        input_size: !ref <text_encoder_input_size>
+        output_size: 1024
+        attention_heads: 16
+        linear_units: 4096
+        num_blocks: 6
+        dropout_rate: 0.1
+        positional_dropout_rate: 0.1
+        attention_dropout_rate: 0.0
+        normalize_before: True
+        input_layer: 'linear'
+        pos_enc_layer_type: 'rel_pos_espnet'
+        selfattention_layer_type: 'rel_selfattn'
+        use_cnn_module: False
+        macaron_style: False
+        use_dynamic_chunk: False
+        use_dynamic_left_chunk: False
+        static_chunk_size: 1
+    llm: !new:cosyvoice.transformer.encoder.TransformerEncoder
+        input_size: !ref <llm_input_size>
+        output_size: !ref <llm_output_size>
+        attention_heads: 16
+        linear_units: 4096
+        num_blocks: 14
+        dropout_rate: 0.1
+        positional_dropout_rate: 0.1
+        attention_dropout_rate: 0.0
+        input_layer: 'linear_legacy'
+        pos_enc_layer_type: 'rel_pos_espnet'
+        selfattention_layer_type: 'rel_selfattn'
+        static_chunk_size: 1
+    sampling: !name:cosyvoice.utils.common.ras_sampling
+        top_p: 0.8
+        top_k: 25
+        win_size: 10
+        tau_r: 0.1
+
+flow: !new:cosyvoice.flow.flow.MaskedDiffWithXvec
+    input_size: 512
+    output_size: 80
+    spk_embed_dim: !ref <spk_embed_dim>
+    output_type: 'mel'
+    vocab_size: 4096
+    input_frame_rate: 50 # change to 25 if you want to train with CosyVoice-300M-25Hz recipe
+    only_mask_loss: True
+    encoder: !new:cosyvoice.transformer.encoder.ConformerEncoder
+        output_size: 512
+        attention_heads: 8
+        linear_units: 2048
+        num_blocks: 6
+        dropout_rate: 0.1
+        positional_dropout_rate: 0.1
+        attention_dropout_rate: 0.1
+        normalize_before: True
+        input_layer: 'linear'
+        pos_enc_layer_type: 'rel_pos_espnet'
+        selfattention_layer_type: 'rel_selfattn'
+        input_size: 512
+        use_cnn_module: False
+        macaron_style: False
+    length_regulator: !new:cosyvoice.flow.length_regulator.InterpolateRegulator
+        channels: 80
+        sampling_ratios: [1, 1, 1, 1]
+    decoder: !new:cosyvoice.flow.flow_matching.ConditionalCFM
+        in_channels: 240
+        n_spks: 1
+        spk_emb_dim: 80
+        cfm_params: !new:omegaconf.DictConfig
+            content:
+                sigma_min: 1e-06
+                solver: 'euler'
+                t_scheduler: 'cosine'
+                training_cfg_rate: 0.2
+                inference_cfg_rate: 0.7
+                reg_loss_type: 'l1'
+        estimator: !new:cosyvoice.flow.decoder.ConditionalDecoder
+            in_channels: 320
+            out_channels: 80
+            channels: [256, 256]
+            dropout: 0.0
+            attention_head_dim: 64
+            n_blocks: 4
+            num_mid_blocks: 12
+            num_heads: 8
+            act_fn: 'gelu'
+
+hift: !new:cosyvoice.hifigan.generator.HiFTGenerator
+    in_channels: 80
+    base_channels: 512
+    nb_harmonics: 8
+    sampling_rate: !ref <sample_rate>
+    nsf_alpha: 0.1
+    nsf_sigma: 0.003
+    nsf_voiced_threshold: 10
+    upsample_rates: [8, 8]
+    upsample_kernel_sizes: [16, 16]
+    istft_params:
+        n_fft: 16
+        hop_len: 4
+    resblock_kernel_sizes: [3, 7, 11]
+    resblock_dilation_sizes: [[1, 3, 5], [1, 3, 5], [1, 3, 5]]
+    source_resblock_kernel_sizes: [7, 11]
+    source_resblock_dilation_sizes: [[1, 3, 5], [1, 3, 5]]
+    lrelu_slope: 0.1
+    audio_limit: 0.99
+    f0_predictor: !new:cosyvoice.hifigan.f0_predictor.ConvRNNF0Predictor
+        num_class: 1
+        in_channels: 80
+        cond_channels: 512
+
+# processor functions
+parquet_opener: !name:cosyvoice.dataset.processor.parquet_opener
+get_tokenizer: !name:whisper.tokenizer.get_tokenizer # change to !name:cosyvoice.tokenizer.tokenizer.get_tokenizer if you want to train with CosyVoice-300M-25Hz recipe
+    multilingual: True
+    num_languages: 100
+    language: 'en'
+    task: 'transcribe'
+allowed_special: 'all'
+tokenize: !name:cosyvoice.dataset.processor.tokenize
+    get_tokenizer: !ref <get_tokenizer>
+    allowed_special: !ref <allowed_special>
+filter: !name:cosyvoice.dataset.processor.filter
+    max_length: 40960
+    min_length: 0
+    token_max_length: 200
+    token_min_length: 1
+resample: !name:cosyvoice.dataset.processor.resample
+    resample_rate: !ref <sample_rate>
+feat_extractor: !name:matcha.utils.audio.mel_spectrogram
+    n_fft: 1024
+    num_mels: 80
+    sampling_rate: !ref <sample_rate>
+    hop_size: 256
+    win_size: 1024
+    fmin: 0
+    fmax: 8000
+    center: False
+compute_fbank: !name:cosyvoice.dataset.processor.compute_fbank
+    feat_extractor: !ref <feat_extractor>
+parse_embedding: !name:cosyvoice.dataset.processor.parse_embedding
+    normalize: True
+shuffle: !name:cosyvoice.dataset.processor.shuffle
+    shuffle_size: 1000
+sort: !name:cosyvoice.dataset.processor.sort
+    sort_size: 500  # sort_size should be less than shuffle_size
+batch: !name:cosyvoice.dataset.processor.batch
+    batch_type: 'dynamic'
+    max_frames_in_batch: 2000
+padding: !name:cosyvoice.dataset.processor.padding
+    use_spk_embedding: False # change to True during sft
+
+# dataset processor pipeline
+data_pipeline: [
+    !ref <parquet_opener>,
+    !ref <tokenize>,
+    !ref <filter>,
+    !ref <resample>,
+    !ref <compute_fbank>,
+    !ref <parse_embedding>,
+    !ref <shuffle>,
+    !ref <sort>,
+    !ref <batch>,
+    !ref <padding>,
+]
+
+# train conf
+train_conf:
+    optim: adam
+    optim_conf:
+        lr: 0.001 # change to 1e-5 during sft
+    scheduler: warmuplr # change to constantlr during sft
+    scheduler_conf:
+        warmup_steps: 2500
+    max_epoch: 200
+    grad_clip: 5
+    accum_grad: 2
+    log_interval: 100
+    save_per_step: -1

runtime/python/fastapi/client.py

@@ -0,0 +1,92 @@
+# Copyright (c) 2024 Alibaba Inc (authors: Xiang Lyu)
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#   http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import argparse
+import logging
+import requests
+import torch
+import torchaudio
+import numpy as np
+
+
+def main():
+    url = "http://{}:{}/inference_{}".format(args.host, args.port, args.mode)
+    if args.mode == 'sft':
+        payload = {
+            'tts_text': args.tts_text,
+            'spk_id': args.spk_id
+        }
+        response = requests.request("GET", url, data=payload, stream=True)
+    elif args.mode == 'zero_shot':
+        payload = {
+            'tts_text': args.tts_text,
+            'prompt_text': args.prompt_text
+        }
+        files = [('prompt_wav', ('prompt_wav', open(args.prompt_wav, 'rb'), 'application/octet-stream'))]
+        response = requests.request("GET", url, data=payload, files=files, stream=True)
+    elif args.mode == 'cross_lingual':
+        payload = {
+            'tts_text': args.tts_text,
+        }
+        files = [('prompt_wav', ('prompt_wav', open(args.prompt_wav, 'rb'), 'application/octet-stream'))]
+        response = requests.request("GET", url, data=payload, files=files, stream=True)
+    else:
+        payload = {
+            'tts_text': args.tts_text,
+            'spk_id': args.spk_id,
+            'instruct_text': args.instruct_text
+        }
+        response = requests.request("GET", url, data=payload, stream=True)
+    tts_audio = b''
+    for r in response.iter_content(chunk_size=16000):
+        tts_audio += r
+    tts_speech = torch.from_numpy(np.array(np.frombuffer(tts_audio, dtype=np.int16))).unsqueeze(dim=0)
+    logging.info('save response to {}'.format(args.tts_wav))
+    torchaudio.save(args.tts_wav, tts_speech, target_sr)
+    logging.info('get response')
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--host',
+                        type=str,
+                        default='0.0.0.0')
+    parser.add_argument('--port',
+                        type=int,
+                        default='50000')
+    parser.add_argument('--mode',
+                        default='sft',
+                        choices=['sft', 'zero_shot', 'cross_lingual', 'instruct'],
+                        help='request mode')
+    parser.add_argument('--tts_text',
+                        type=str,
+                        default='你好，我是通义千问语音合成大模型，请问有什么可以帮您的吗？')
+    parser.add_argument('--spk_id',
+                        type=str,
+                        default='中文女')
+    parser.add_argument('--prompt_text',
+                        type=str,
+                        default='希望你以后能够做的比我还好呦。')
+    parser.add_argument('--prompt_wav',
+                        type=str,
+                        default='../../../asset/zero_shot_prompt.wav')
+    parser.add_argument('--instruct_text',
+                        type=str,
+                        default='Theo \'Crimson\', is a fiery, passionate rebel leader. \
+                                 Fights with fervor for justice, but struggles with impulsiveness.')
+    parser.add_argument('--tts_wav',
+                        type=str,
+                        default='demo.wav')
+    args = parser.parse_args()
+    prompt_sr, target_sr = 16000, 22050
+    main()

third_party/Matcha-TTS/.github/PULL_REQUEST_TEMPLATE.md

@@ -0,0 +1,22 @@
+## What does this PR do?
+
+<!--
+Please include a summary of the change and which issue is fixed.
+Please also include relevant motivation and context.
+List any dependencies that are required for this change.
+List all the breaking changes introduced by this pull request.
+-->
+
+Fixes #\<issue_number>
+
+## Before submitting
+
+- [ ] Did you make sure **title is self-explanatory** and **the description concisely explains the PR**?
+- [ ] Did you make sure your **PR does only one thing**, instead of bundling different changes together?
+- [ ] Did you list all the **breaking changes** introduced by this pull request?
+- [ ] Did you **test your PR locally** with `pytest` command?
+- [ ] Did you **run pre-commit hooks** with `pre-commit run -a` command?
+
+## Did you have fun?
+
+Make sure you had fun coding 🙃

third_party/Matcha-TTS/.github/dependabot.yml

@@ -0,0 +1,17 @@
+# To get started with Dependabot version updates, you'll need to specify which
+# package ecosystems to update and where the package manifests are located.
+# Please see the documentation for all configuration options:
+# https://docs.github.com/github/administering-a-repository/configuration-options-for-dependency-updates
+
+version: 2
+updates:
+  - package-ecosystem: "pip" # See documentation for possible values
+    directory: "/" # Location of package manifests
+    target-branch: "dev"
+    schedule:
+      interval: "daily"
+    ignore:
+      - dependency-name: "pytorch-lightning"
+        update-types: ["version-update:semver-patch"]
+      - dependency-name: "torchmetrics"
+        update-types: ["version-update:semver-patch"]

third_party/Matcha-TTS/.github/release-drafter.yml

@@ -0,0 +1,44 @@
+name-template: "v$RESOLVED_VERSION"
+tag-template: "v$RESOLVED_VERSION"
+
+categories:
+  - title: "🚀 Features"
+    labels:
+      - "feature"
+      - "enhancement"
+  - title: "🐛 Bug Fixes"
+    labels:
+      - "fix"
+      - "bugfix"
+      - "bug"
+  - title: "🧹 Maintenance"
+    labels:
+      - "maintenance"
+      - "dependencies"
+      - "refactoring"
+      - "cosmetic"
+      - "chore"
+  - title: "📝️ Documentation"
+    labels:
+      - "documentation"
+      - "docs"
+
+change-template: "- $TITLE @$AUTHOR (#$NUMBER)"
+change-title-escapes: '\<*_&' # You can add # and @ to disable mentions
+
+version-resolver:
+  major:
+    labels:
+      - "major"
+  minor:
+    labels:
+      - "minor"
+  patch:
+    labels:
+      - "patch"
+  default: patch
+
+template: |
+  ## Changes
+
+  $CHANGES

third_party/Matcha-TTS/.pylintrc

@@ -0,0 +1,525 @@
+[MASTER]
+
+# A comma-separated list of package or module names from where C extensions may
+# be loaded. Extensions are loading into the active Python interpreter and may
+# run arbitrary code.
+extension-pkg-whitelist=
+
+# Add files or directories to the blacklist. They should be base names, not
+# paths.
+ignore=CVS
+
+# Add files or directories matching the regex patterns to the blacklist. The
+# regex matches against base names, not paths.
+ignore-patterns=
+
+# Python code to execute, usually for sys.path manipulation such as
+# pygtk.require().
+#init-hook=
+
+# Use multiple processes to speed up Pylint. Specifying 0 will auto-detect the
+# number of processors available to use.
+jobs=1
+
+# Control the amount of potential inferred values when inferring a single
+# object. This can help the performance when dealing with large functions or
+# complex, nested conditions.
+limit-inference-results=100
+
+# List of plugins (as comma separated values of python modules names) to load,
+# usually to register additional checkers.
+load-plugins=
+
+# Pickle collected data for later comparisons.
+persistent=yes
+
+# Specify a configuration file.
+#rcfile=
+
+# When enabled, pylint would attempt to guess common misconfiguration and emit
+# user-friendly hints instead of false-positive error messages.
+suggestion-mode=yes
+
+# Allow loading of arbitrary C extensions. Extensions are imported into the
+# active Python interpreter and may run arbitrary code.
+unsafe-load-any-extension=no
+
+
+[MESSAGES CONTROL]
+
+# Only show warnings with the listed confidence levels. Leave empty to show
+# all. Valid levels: HIGH, INFERENCE, INFERENCE_FAILURE, UNDEFINED.
+confidence=
+
+# Disable the message, report, category or checker with the given id(s). You
+# can either give multiple identifiers separated by comma (,) or put this
+# option multiple times (only on the command line, not in the configuration
+# file where it should appear only once). You can also use "--disable=all" to
+# disable everything first and then reenable specific checks. For example, if
+# you want to run only the similarities checker, you can use "--disable=all
+# --enable=similarities". If you want to run only the classes checker, but have
+# no Warning level messages displayed, use "--disable=all --enable=classes
+# --disable=W".
+disable=missing-docstring,
+        too-many-public-methods,
+        too-many-lines,
+        bare-except,
+        ## for avoiding weird p3.6 CI linter error
+        ## TODO: see later if we can remove this
+        assigning-non-slot,
+        unsupported-assignment-operation,
+        ## end
+        line-too-long,
+        fixme,
+        wrong-import-order,
+        ungrouped-imports,
+        wrong-import-position,
+        import-error,
+        invalid-name,
+        too-many-instance-attributes,
+        arguments-differ,
+        arguments-renamed,
+        no-name-in-module,
+        no-member,
+        unsubscriptable-object,
+        raw-checker-failed,
+        bad-inline-option,
+        locally-disabled,
+        file-ignored,
+        suppressed-message,
+        useless-suppression,
+        deprecated-pragma,
+        use-symbolic-message-instead,
+        useless-object-inheritance,
+        too-few-public-methods,
+        too-many-branches,
+        too-many-arguments,
+        too-many-locals,
+        too-many-statements,
+        duplicate-code,
+        not-callable,
+        import-outside-toplevel,
+        logging-fstring-interpolation,
+        logging-not-lazy,
+        unused-argument,
+        no-else-return,
+        chained-comparison,
+        redefined-outer-name
+
+# Enable the message, report, category or checker with the given id(s). You can
+# either give multiple identifier separated by comma (,) or put this option
+# multiple time (only on the command line, not in the configuration file where
+# it should appear only once). See also the "--disable" option for examples.
+enable=c-extension-no-member
+
+
+[REPORTS]
+
+# Python expression which should return a note less than 10 (10 is the highest
+# note). You have access to the variables errors warning, statement which
+# respectively contain the number of errors / warnings messages and the total
+# number of statements analyzed. This is used by the global evaluation report
+# (RP0004).
+evaluation=10.0 - ((float(5 * error + warning + refactor + convention) / statement) * 10)
+
+# Template used to display messages. This is a python new-style format string
+# used to format the message information. See doc for all details.
+#msg-template=
+
+# Set the output format. Available formats are text, parseable, colorized, json
+# and msvs (visual studio). You can also give a reporter class, e.g.
+# mypackage.mymodule.MyReporterClass.
+output-format=text
+
+# Tells whether to display a full report or only the messages.
+reports=no
+
+# Activate the evaluation score.
+score=yes
+
+
+[REFACTORING]
+
+# Maximum number of nested blocks for function / method body
+max-nested-blocks=5
+
+# Complete name of functions that never returns. When checking for
+# inconsistent-return-statements if a never returning function is called then
+# it will be considered as an explicit return statement and no message will be
+# printed.
+never-returning-functions=sys.exit
+
+
+[LOGGING]
+
+# Format style used to check logging format string. `old` means using %
+# formatting, while `new` is for `{}` formatting.
+logging-format-style=old
+
+# Logging modules to check that the string format arguments are in logging
+# function parameter format.
+logging-modules=logging
+
+
+[SPELLING]
+
+# Limits count of emitted suggestions for spelling mistakes.
+max-spelling-suggestions=4
+
+# Spelling dictionary name. Available dictionaries: none. To make it working
+# install python-enchant package..
+spelling-dict=
+
+# List of comma separated words that should not be checked.
+spelling-ignore-words=
+
+# A path to a file that contains private dictionary; one word per line.
+spelling-private-dict-file=
+
+# Tells whether to store unknown words to indicated private dictionary in
+# --spelling-private-dict-file option instead of raising a message.
+spelling-store-unknown-words=no
+
+
+[MISCELLANEOUS]
+
+# List of note tags to take in consideration, separated by a comma.
+notes=FIXME,
+      XXX,
+      TODO
+
+
+[TYPECHECK]
+
+# List of decorators that produce context managers, such as
+# contextlib.contextmanager. Add to this list to register other decorators that
+# produce valid context managers.
+contextmanager-decorators=contextlib.contextmanager
+
+# List of members which are set dynamically and missed by pylint inference
+# system, and so shouldn't trigger E1101 when accessed. Python regular
+# expressions are accepted.
+generated-members=numpy.*,torch.*
+
+# Tells whether missing members accessed in mixin class should be ignored. A
+# mixin class is detected if its name ends with "mixin" (case insensitive).
+ignore-mixin-members=yes
+
+# Tells whether to warn about missing members when the owner of the attribute
+# is inferred to be None.
+ignore-none=yes
+
+# This flag controls whether pylint should warn about no-member and similar
+# checks whenever an opaque object is returned when inferring. The inference
+# can return multiple potential results while evaluating a Python object, but
+# some branches might not be evaluated, which results in partial inference. In
+# that case, it might be useful to still emit no-member and other checks for
+# the rest of the inferred objects.
+ignore-on-opaque-inference=yes
+
+# List of class names for which member attributes should not be checked (useful
+# for classes with dynamically set attributes). This supports the use of
+# qualified names.
+ignored-classes=optparse.Values,thread._local,_thread._local
+
+# List of module names for which member attributes should not be checked
+# (useful for modules/projects where namespaces are manipulated during runtime
+# and thus existing member attributes cannot be deduced by static analysis. It
+# supports qualified module names, as well as Unix pattern matching.
+ignored-modules=
+
+# Show a hint with possible names when a member name was not found. The aspect
+# of finding the hint is based on edit distance.
+missing-member-hint=yes
+
+# The minimum edit distance a name should have in order to be considered a
+# similar match for a missing member name.
+missing-member-hint-distance=1
+
+# The total number of similar names that should be taken in consideration when
+# showing a hint for a missing member.
+missing-member-max-choices=1
+
+
+[VARIABLES]
+
+# List of additional names supposed to be defined in builtins. Remember that
+# you should avoid defining new builtins when possible.
+additional-builtins=
+
+# Tells whether unused global variables should be treated as a violation.
+allow-global-unused-variables=yes
+
+# List of strings which can identify a callback function by name. A callback
+# name must start or end with one of those strings.
+callbacks=cb_,
+          _cb
+
+# A regular expression matching the name of dummy variables (i.e. expected to
+# not be used).
+dummy-variables-rgx=_+$|(_[a-zA-Z0-9_]*[a-zA-Z0-9]+?$)|dummy|^ignored_|^unused_
+
+# Argument names that match this expression will be ignored. Default to name
+# with leading underscore.
+ignored-argument-names=_.*|^ignored_|^unused_
+
+# Tells whether we should check for unused import in __init__ files.
+init-import=no
+
+# List of qualified module names which can have objects that can redefine
+# builtins.
+redefining-builtins-modules=six.moves,past.builtins,future.builtins,builtins,io
+
+
+[FORMAT]
+
+# Expected format of line ending, e.g. empty (any line ending), LF or CRLF.
+expected-line-ending-format=
+
+# Regexp for a line that is allowed to be longer than the limit.
+ignore-long-lines=^\s*(# )?<?https?://\S+>?$
+
+# Number of spaces of indent required inside a hanging or continued line.
+indent-after-paren=4
+
+# String used as indentation unit. This is usually "    " (4 spaces) or "\t" (1
+# tab).
+indent-string='    '
+
+# Maximum number of characters on a single line.
+max-line-length=120
+
+# Maximum number of lines in a module.
+max-module-lines=1000
+
+# Allow the body of a class to be on the same line as the declaration if body
+# contains single statement.
+single-line-class-stmt=no
+
+# Allow the body of an if to be on the same line as the test if there is no
+# else.
+single-line-if-stmt=no
+
+
+[SIMILARITIES]
+
+# Ignore comments when computing similarities.
+ignore-comments=yes
+
+# Ignore docstrings when computing similarities.
+ignore-docstrings=yes
+
+# Ignore imports when computing similarities.
+ignore-imports=no
+
+# Minimum lines number of a similarity.
+min-similarity-lines=4
+
+
+[BASIC]
+
+# Naming style matching correct argument names.
+argument-naming-style=snake_case
+
+# Regular expression matching correct argument names. Overrides argument-
+# naming-style.
+argument-rgx=[a-z_][a-z0-9_]{0,30}$
+
+# Naming style matching correct attribute names.
+attr-naming-style=snake_case
+
+# Regular expression matching correct attribute names. Overrides attr-naming-
+# style.
+#attr-rgx=
+
+# Bad variable names which should always be refused, separated by a comma.
+bad-names=
+
+# Naming style matching correct class attribute names.
+class-attribute-naming-style=any
+
+# Regular expression matching correct class attribute names. Overrides class-
+# attribute-naming-style.
+#class-attribute-rgx=
+
+# Naming style matching correct class names.
+class-naming-style=PascalCase
+
+# Regular expression matching correct class names. Overrides class-naming-
+# style.
+#class-rgx=
+
+# Naming style matching correct constant names.
+const-naming-style=UPPER_CASE
+
+# Regular expression matching correct constant names. Overrides const-naming-
+# style.
+#const-rgx=
+
+# Minimum line length for functions/classes that require docstrings, shorter
+# ones are exempt.
+docstring-min-length=-1
+
+# Naming style matching correct function names.
+function-naming-style=snake_case
+
+# Regular expression matching correct function names. Overrides function-
+# naming-style.
+#function-rgx=
+
+# Good variable names which should always be accepted, separated by a comma.
+good-names=i,
+           j,
+           k,
+           x,
+           ex,
+           Run,
+           _
+
+# Include a hint for the correct naming format with invalid-name.
+include-naming-hint=no
+
+# Naming style matching correct inline iteration names.
+inlinevar-naming-style=any
+
+# Regular expression matching correct inline iteration names. Overrides
+# inlinevar-naming-style.
+#inlinevar-rgx=
+
+# Naming style matching correct method names.
+method-naming-style=snake_case
+
+# Regular expression matching correct method names. Overrides method-naming-
+# style.
+#method-rgx=
+
+# Naming style matching correct module names.
+module-naming-style=snake_case
+
+# Regular expression matching correct module names. Overrides module-naming-
+# style.
+#module-rgx=
+
+# Colon-delimited sets of names that determine each other's naming style when
+# the name regexes allow several styles.
+name-group=
+
+# Regular expression which should only match function or class names that do
+# not require a docstring.
+no-docstring-rgx=^_
+
+# List of decorators that produce properties, such as abc.abstractproperty. Add
+# to this list to register other decorators that produce valid properties.
+# These decorators are taken in consideration only for invalid-name.
+property-classes=abc.abstractproperty
+
+# Naming style matching correct variable names.
+variable-naming-style=snake_case
+
+# Regular expression matching correct variable names. Overrides variable-
+# naming-style.
+variable-rgx=[a-z_][a-z0-9_]{0,30}$
+
+
+[STRING]
+
+# This flag controls whether the implicit-str-concat-in-sequence should
+# generate a warning on implicit string concatenation in sequences defined over
+# several lines.
+check-str-concat-over-line-jumps=no
+
+
+[IMPORTS]
+
+# Allow wildcard imports from modules that define __all__.
+allow-wildcard-with-all=no
+
+# Analyse import fallback blocks. This can be used to support both Python 2 and
+# 3 compatible code, which means that the block might have code that exists
+# only in one or another interpreter, leading to false positives when analysed.
+analyse-fallback-blocks=no
+
+# Deprecated modules which should not be used, separated by a comma.
+deprecated-modules=optparse,tkinter.tix
+
+# Create a graph of external dependencies in the given file (report RP0402 must
+# not be disabled).
+ext-import-graph=
+
+# Create a graph of every (i.e. internal and external) dependencies in the
+# given file (report RP0402 must not be disabled).
+import-graph=
+
+# Create a graph of internal dependencies in the given file (report RP0402 must
+# not be disabled).
+int-import-graph=
+
+# Force import order to recognize a module as part of the standard
+# compatibility libraries.
+known-standard-library=
+
+# Force import order to recognize a module as part of a third party library.
+known-third-party=enchant
+
+
+[CLASSES]
+
+# List of method names used to declare (i.e. assign) instance attributes.
+defining-attr-methods=__init__,
+                      __new__,
+                      setUp
+
+# List of member names, which should be excluded from the protected access
+# warning.
+exclude-protected=_asdict,
+                  _fields,
+                  _replace,
+                  _source,
+                  _make
+
+# List of valid names for the first argument in a class method.
+valid-classmethod-first-arg=cls
+
+# List of valid names for the first argument in a metaclass class method.
+valid-metaclass-classmethod-first-arg=cls
+
+
+[DESIGN]
+
+# Maximum number of arguments for function / method.
+max-args=5
+
+# Maximum number of attributes for a class (see R0902).
+max-attributes=7
+
+# Maximum number of boolean expressions in an if statement.
+max-bool-expr=5
+
+# Maximum number of branch for function / method body.
+max-branches=12
+
+# Maximum number of locals for function / method body.
+max-locals=15
+
+# Maximum number of parents for a class (see R0901).
+max-parents=15
+
+# Maximum number of public methods for a class (see R0904).
+max-public-methods=20
+
+# Maximum number of return / yield for function / method body.
+max-returns=6
+
+# Maximum number of statements in function / method body.
+max-statements=50
+
+# Minimum number of public methods for a class (see R0903).
+min-public-methods=2
+
+
+[EXCEPTIONS]
+
+# Exceptions that will emit a warning when being caught. Defaults to
+# "BaseException, Exception".
+overgeneral-exceptions=builtins.BaseException,
+                       builtins.Exception

third_party/Matcha-TTS/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2023 Shivam Mehta
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

third_party/Matcha-TTS/MANIFEST.in

@@ -0,0 +1,14 @@
+include README.md
+include LICENSE.txt
+include requirements.*.txt
+include *.cff
+include requirements.txt
+include matcha/VERSION
+recursive-include matcha *.json
+recursive-include matcha *.html
+recursive-include matcha *.png
+recursive-include matcha *.md
+recursive-include matcha *.py
+recursive-include matcha *.pyx
+recursive-exclude tests *
+prune tests*

third_party/Matcha-TTS/README.md

@@ -0,0 +1,278 @@
+<div align="center">
+
+# 🍵 Matcha-TTS: A fast TTS architecture with conditional flow matching
+
+### [Shivam Mehta](https://www.kth.se/profile/smehta), [Ruibo Tu](https://www.kth.se/profile/ruibo), [Jonas Beskow](https://www.kth.se/profile/beskow), [Éva Székely](https://www.kth.se/profile/szekely), and [Gustav Eje Henter](https://people.kth.se/~ghe/)
+
+[![python](https://img.shields.io/badge/-Python_3.10-blue?logo=python&logoColor=white)](https://www.python.org/downloads/release/python-3100/)
+[![pytorch](https://img.shields.io/badge/PyTorch_2.0+-ee4c2c?logo=pytorch&logoColor=white)](https://pytorch.org/get-started/locally/)
+[![lightning](https://img.shields.io/badge/-Lightning_2.0+-792ee5?logo=pytorchlightning&logoColor=white)](https://pytorchlightning.ai/)
+[![hydra](https://img.shields.io/badge/Config-Hydra_1.3-89b8cd)](https://hydra.cc/)
+[![black](https://img.shields.io/badge/Code%20Style-Black-black.svg?labelColor=gray)](https://black.readthedocs.io/en/stable/)
+[![isort](https://img.shields.io/badge/%20imports-isort-%231674b1?style=flat&labelColor=ef8336)](https://pycqa.github.io/isort/)
+
+<p style="text-align: center;">
+  <img src="https://shivammehta25.github.io/Matcha-TTS/images/logo.png" height="128"/>
+</p>
+
+</div>
+
+> This is the official code implementation of 🍵 Matcha-TTS [ICASSP 2024].
+
+We propose 🍵 Matcha-TTS, a new approach to non-autoregressive neural TTS, that uses [conditional flow matching](https://arxiv.org/abs/2210.02747) (similar to [rectified flows](https://arxiv.org/abs/2209.03003)) to speed up ODE-based speech synthesis. Our method:
+
+- Is probabilistic
+- Has compact memory footprint
+- Sounds highly natural
+- Is very fast to synthesise from
+
+Check out our [demo page](https://shivammehta25.github.io/Matcha-TTS) and read [our ICASSP 2024 paper](https://arxiv.org/abs/2309.03199) for more details.
+
+[Pre-trained models](https://drive.google.com/drive/folders/17C_gYgEHOxI5ZypcfE_k1piKCtyR0isJ?usp=sharing) will be automatically downloaded with the CLI or gradio interface.
+
+You can also [try 🍵 Matcha-TTS in your browser on HuggingFace 🤗 spaces](https://huggingface.co/spaces/shivammehta25/Matcha-TTS).
+
+## Teaser video
+
+[![Watch the video](https://img.youtube.com/vi/xmvJkz3bqw0/hqdefault.jpg)](https://youtu.be/xmvJkz3bqw0)
+
+## Installation
+
+1. Create an environment (suggested but optional)
+
+```
+conda create -n matcha-tts python=3.10 -y
+conda activate matcha-tts
+```
+
+2. Install Matcha TTS using pip or from source
+
+```bash
+pip install matcha-tts
+```
+
+from source
+
+```bash
+pip install git+https://github.com/shivammehta25/Matcha-TTS.git
+cd Matcha-TTS
+pip install -e .
+```
+
+3. Run CLI / gradio app / jupyter notebook
+
+```bash
+# This will download the required models
+matcha-tts --text "<INPUT TEXT>"
+```
+
+or
+
+```bash
+matcha-tts-app
+```
+
+or open `synthesis.ipynb` on jupyter notebook
+
+### CLI Arguments
+
+- To synthesise from given text, run:
+
+```bash
+matcha-tts --text "<INPUT TEXT>"
+```
+
+- To synthesise from a file, run:
+
+```bash
+matcha-tts --file <PATH TO FILE>
+```
+
+- To batch synthesise from a file, run:
+
+```bash
+matcha-tts --file <PATH TO FILE> --batched
+```
+
+Additional arguments
+
+- Speaking rate
+
+```bash
+matcha-tts --text "<INPUT TEXT>" --speaking_rate 1.0
+```
+
+- Sampling temperature
+
+```bash
+matcha-tts --text "<INPUT TEXT>" --temperature 0.667
+```
+
+- Euler ODE solver steps
+
+```bash
+matcha-tts --text "<INPUT TEXT>" --steps 10
+```
+
+## Train with your own dataset
+
+Let's assume we are training with LJ Speech
+
+1. Download the dataset from [here](https://keithito.com/LJ-Speech-Dataset/), extract it to `data/LJSpeech-1.1`, and prepare the file lists to point to the extracted data like for [item 5 in the setup of the NVIDIA Tacotron 2 repo](https://github.com/NVIDIA/tacotron2#setup).
+
+2. Clone and enter the Matcha-TTS repository
+
+```bash
+git clone https://github.com/shivammehta25/Matcha-TTS.git
+cd Matcha-TTS
+```
+
+3. Install the package from source
+
+```bash
+pip install -e .
+```
+
+4. Go to `configs/data/ljspeech.yaml` and change
+
+```yaml
+train_filelist_path: data/filelists/ljs_audio_text_train_filelist.txt
+valid_filelist_path: data/filelists/ljs_audio_text_val_filelist.txt
+```
+
+5. Generate normalisation statistics with the yaml file of dataset configuration
+
+```bash
+matcha-data-stats -i ljspeech.yaml
+# Output:
+#{'mel_mean': -5.53662231756592, 'mel_std': 2.1161014277038574}
+```
+
+Update these values in `configs/data/ljspeech.yaml` under `data_statistics` key.
+
+```bash
+data_statistics:  # Computed for ljspeech dataset
+  mel_mean: -5.536622
+  mel_std: 2.116101
+```
+
+to the paths of your train and validation filelists.
+
+6. Run the training script
+
+```bash
+make train-ljspeech
+```
+
+or
+
+```bash
+python matcha/train.py experiment=ljspeech
+```
+
+- for a minimum memory run
+
+```bash
+python matcha/train.py experiment=ljspeech_min_memory
+```
+
+- for multi-gpu training, run
+
+```bash
+python matcha/train.py experiment=ljspeech trainer.devices=[0,1]
+```
+
+7. Synthesise from the custom trained model
+
+```bash
+matcha-tts --text "<INPUT TEXT>" --checkpoint_path <PATH TO CHECKPOINT>
+```
+
+## ONNX support
+
+> Special thanks to [@mush42](https://github.com/mush42) for implementing ONNX export and inference support.
+
+It is possible to export Matcha checkpoints to [ONNX](https://onnx.ai/), and run inference on the exported ONNX graph.
+
+### ONNX export
+
+To export a checkpoint to ONNX, first install ONNX with
+
+```bash
+pip install onnx
+```
+
+then run the following:
+
+```bash
+python3 -m matcha.onnx.export matcha.ckpt model.onnx --n-timesteps 5
+```
+
+Optionally, the ONNX exporter accepts **vocoder-name** and **vocoder-checkpoint** arguments. This enables you to embed the vocoder in the exported graph and generate waveforms in a single run (similar to end-to-end TTS systems).
+
+**Note** that `n_timesteps` is treated as a hyper-parameter rather than a model input. This means you should specify it during export (not during inference). If not specified, `n_timesteps` is set to **5**.
+
+**Important**: for now, torch>=2.1.0 is needed for export since the `scaled_product_attention` operator is not exportable in older versions. Until the final version is released, those who want to export their models must install torch>=2.1.0 manually as a pre-release.
+
+### ONNX Inference
+
+To run inference on the exported model, first install `onnxruntime` using
+
+```bash
+pip install onnxruntime
+pip install onnxruntime-gpu  # for GPU inference
+```
+
+then use the following:
+
+```bash
+python3 -m matcha.onnx.infer model.onnx --text "hey" --output-dir ./outputs
+```
+
+You can also control synthesis parameters:
+
+```bash
+python3 -m matcha.onnx.infer model.onnx --text "hey" --output-dir ./outputs --temperature 0.4 --speaking_rate 0.9 --spk 0
+```
+
+To run inference on **GPU**, make sure to install **onnxruntime-gpu** package, and then pass `--gpu` to the inference command:
+
+```bash
+python3 -m matcha.onnx.infer model.onnx --text "hey" --output-dir ./outputs --gpu
+```
+
+If you exported only Matcha to ONNX, this will write mel-spectrogram as graphs and `numpy` arrays to the output directory.
+If you embedded the vocoder in the exported graph, this will write `.wav` audio files to the output directory.
+
+If you exported only Matcha to ONNX, and you want to run a full TTS pipeline, you can pass a path to a vocoder model in `ONNX` format:
+
+```bash
+python3 -m matcha.onnx.infer model.onnx --text "hey" --output-dir ./outputs --vocoder hifigan.small.onnx
+```
+
+This will write `.wav` audio files to the output directory.
+
+## Citation information
+
+If you use our code or otherwise find this work useful, please cite our paper:
+
+```text
+@inproceedings{mehta2024matcha,
+  title={Matcha-{TTS}: A fast {TTS} architecture with conditional flow matching},
+  author={Mehta, Shivam and Tu, Ruibo and Beskow, Jonas and Sz{\'e}kely, {\'E}va and Henter, Gustav Eje},
+  booktitle={Proc. ICASSP},
+  year={2024}
+}
+```
+
+## Acknowledgements
+
+Since this code uses [Lightning-Hydra-Template](https://github.com/ashleve/lightning-hydra-template), you have all the powers that come with it.
+
+Other source code we would like to acknowledge:
+
+- [Coqui-TTS](https://github.com/coqui-ai/TTS/tree/dev): For helping me figure out how to make cython binaries pip installable and encouragement
+- [Hugging Face Diffusers](https://huggingface.co/): For their awesome diffusers library and its components
+- [Grad-TTS](https://github.com/huawei-noah/Speech-Backbones/tree/main/Grad-TTS): For the monotonic alignment search source code
+- [torchdyn](https://github.com/DiffEqML/torchdyn): Useful for trying other ODE solvers during research and development
+- [labml.ai](https://nn.labml.ai/transformers/rope/index.html): For the RoPE implementation

third_party/Matcha-TTS/configs/callbacks/model_checkpoint.yaml

@@ -0,0 +1,17 @@
+# https://lightning.ai/docs/pytorch/stable/api/lightning.pytorch.callbacks.ModelCheckpoint.html
+
+model_checkpoint:
+  _target_: lightning.pytorch.callbacks.ModelCheckpoint
+  dirpath: ${paths.output_dir}/checkpoints # directory to save the model file
+  filename: checkpoint_{epoch:03d}  # checkpoint filename
+  monitor: epoch # name of the logged metric which determines when model is improving
+  verbose: False # verbosity mode
+  save_last: true # additionally always save an exact copy of the last checkpoint to a file last.ckpt
+  save_top_k: 10 # save k best models (determined by above metric)
+  mode: "max" # "max" means higher metric value is better, can be also "min"
+  auto_insert_metric_name: True # when True, the checkpoints filenames will contain the metric name
+  save_weights_only: False # if True, then only the model’s weights will be saved
+  every_n_train_steps: null # number of training steps between checkpoints
+  train_time_interval: null # checkpoints are monitored at the specified time interval
+  every_n_epochs: 100 # number of epochs between checkpoints
+  save_on_train_epoch_end: null # whether to run checkpointing at the end of the training epoch or the end of validation

third_party/Matcha-TTS/configs/callbacks/rich_progress_bar.yaml

@@ -0,0 +1,4 @@
+# https://lightning.ai/docs/pytorch/latest/api/lightning.pytorch.callbacks.RichProgressBar.html
+
+rich_progress_bar:
+  _target_: lightning.pytorch.callbacks.RichProgressBar

third_party/Matcha-TTS/configs/debug/default.yaml

@@ -0,0 +1,35 @@
+# @package _global_
+
+# default debugging setup, runs 1 full epoch
+# other debugging configs can inherit from this one
+
+# overwrite task name so debugging logs are stored in separate folder
+task_name: "debug"
+
+# disable callbacks and loggers during debugging
+# callbacks: null
+# logger: null
+
+extras:
+  ignore_warnings: False
+  enforce_tags: False
+
+# sets level of all command line loggers to 'DEBUG'
+# https://hydra.cc/docs/tutorials/basic/running_your_app/logging/
+hydra:
+  job_logging:
+    root:
+      level: DEBUG
+
+  # use this to also set hydra loggers to 'DEBUG'
+  # verbose: True
+
+trainer:
+  max_epochs: 1
+  accelerator: cpu # debuggers don't like gpus
+  devices: 1 # debuggers don't like multiprocessing
+  detect_anomaly: true # raise exception if NaN or +/-inf is detected in any tensor
+
+data:
+  num_workers: 0 # debuggers don't like multiprocessing
+  pin_memory: False # disable gpu memory pin

third_party/Matcha-TTS/configs/debug/overfit.yaml

@@ -0,0 +1,13 @@
+# @package _global_
+
+# overfits to 3 batches
+
+defaults:
+  - default
+
+trainer:
+  max_epochs: 20
+  overfit_batches: 3
+
+# model ckpt and early stopping need to be disabled during overfitting
+callbacks: null

third_party/Matcha-TTS/configs/debug/profiler.yaml

@@ -0,0 +1,15 @@
+# @package _global_
+
+# runs with execution time profiling
+
+defaults:
+  - default
+
+trainer:
+  max_epochs: 1
+  # profiler: "simple"
+  profiler: "advanced"
+  # profiler: "pytorch"
+  accelerator: gpu
+
+  limit_train_batches: 0.02

third_party/Matcha-TTS/configs/eval.yaml

@@ -0,0 +1,18 @@
+# @package _global_
+
+defaults:
+  - _self_
+  - data: mnist # choose datamodule with `test_dataloader()` for evaluation
+  - model: mnist
+  - logger: null
+  - trainer: default
+  - paths: default
+  - extras: default
+  - hydra: default
+
+task_name: "eval"
+
+tags: ["dev"]
+
+# passing checkpoint path is necessary for evaluation
+ckpt_path: ???

third_party/Matcha-TTS/configs/experiment/ljspeech.yaml

@@ -0,0 +1,14 @@
+# @package _global_
+
+# to execute this experiment run:
+# python train.py experiment=multispeaker
+
+defaults:
+  - override /data: ljspeech.yaml
+
+# all parameters below will be merged with parameters from default configurations set above
+# this allows you to overwrite only specified parameters
+
+tags: ["ljspeech"]
+
+run_name: ljspeech

third_party/Matcha-TTS/configs/experiment/ljspeech_min_memory.yaml

@@ -0,0 +1,18 @@
+# @package _global_
+
+# to execute this experiment run:
+# python train.py experiment=multispeaker
+
+defaults:
+  - override /data: ljspeech.yaml
+
+# all parameters below will be merged with parameters from default configurations set above
+# this allows you to overwrite only specified parameters
+
+tags: ["ljspeech"]
+
+run_name: ljspeech_min
+
+
+model:
+  out_size: 172

third_party/Matcha-TTS/configs/experiment/multispeaker.yaml

@@ -0,0 +1,14 @@
+# @package _global_
+
+# to execute this experiment run:
+# python train.py experiment=multispeaker
+
+defaults:
+  - override /data: vctk.yaml
+
+# all parameters below will be merged with parameters from default configurations set above
+# this allows you to overwrite only specified parameters
+
+tags: ["multispeaker"]
+
+run_name: multispeaker

third_party/Matcha-TTS/configs/extras/default.yaml

@@ -0,0 +1,8 @@
+# disable python warnings if they annoy you
+ignore_warnings: False
+
+# ask user for tags if none are provided in the config
+enforce_tags: True
+
+# pretty print config tree at the start of the run using Rich library
+print_config: True

third_party/Matcha-TTS/configs/hparams_search/mnist_optuna.yaml

@@ -0,0 +1,52 @@
+# @package _global_
+
+# example hyperparameter optimization of some experiment with Optuna:
+# python train.py -m hparams_search=mnist_optuna experiment=example
+
+defaults:
+  - override /hydra/sweeper: optuna
+
+# choose metric which will be optimized by Optuna
+# make sure this is the correct name of some metric logged in lightning module!
+optimized_metric: "val/acc_best"
+
+# here we define Optuna hyperparameter search
+# it optimizes for value returned from function with @hydra.main decorator
+# docs: https://hydra.cc/docs/next/plugins/optuna_sweeper
+hydra:
+  mode: "MULTIRUN" # set hydra to multirun by default if this config is attached
+
+  sweeper:
+    _target_: hydra_plugins.hydra_optuna_sweeper.optuna_sweeper.OptunaSweeper
+
+    # storage URL to persist optimization results
+    # for example, you can use SQLite if you set 'sqlite:///example.db'
+    storage: null
+
+    # name of the study to persist optimization results
+    study_name: null
+
+    # number of parallel workers
+    n_jobs: 1
+
+    # 'minimize' or 'maximize' the objective
+    direction: maximize
+
+    # total number of runs that will be executed
+    n_trials: 20
+
+    # choose Optuna hyperparameter sampler
+    # you can choose bayesian sampler (tpe), random search (without optimization), grid sampler, and others
+    # docs: https://optuna.readthedocs.io/en/stable/reference/samplers.html
+    sampler:
+      _target_: optuna.samplers.TPESampler
+      seed: 1234
+      n_startup_trials: 10 # number of random sampling runs before optimization starts
+
+    # define hyperparameter search space
+    params:
+      model.optimizer.lr: interval(0.0001, 0.1)
+      data.batch_size: choice(32, 64, 128, 256)
+      model.net.lin1_size: choice(64, 128, 256)
+      model.net.lin2_size: choice(64, 128, 256)
+      model.net.lin3_size: choice(32, 64, 128, 256)

third_party/Matcha-TTS/configs/logger/aim.yaml

@@ -0,0 +1,28 @@
+# https://aimstack.io/
+
+# example usage in lightning module:
+# https://github.com/aimhubio/aim/blob/main/examples/pytorch_lightning_track.py
+
+# open the Aim UI with the following command (run in the folder containing the `.aim` folder):
+# `aim up`
+
+aim:
+  _target_: aim.pytorch_lightning.AimLogger
+  repo: ${paths.root_dir} # .aim folder will be created here
+  # repo: "aim://ip_address:port" # can instead provide IP address pointing to Aim remote tracking server which manages the repo, see https://aimstack.readthedocs.io/en/latest/using/remote_tracking.html#
+
+  # aim allows to group runs under experiment name
+  experiment: null # any string, set to "default" if not specified
+
+  train_metric_prefix: "train/"
+  val_metric_prefix: "val/"
+  test_metric_prefix: "test/"
+
+  # sets the tracking interval in seconds for system usage metrics (CPU, GPU, memory, etc.)
+  system_tracking_interval: 10 # set to null to disable system metrics tracking
+
+  # enable/disable logging of system params such as installed packages, git info, env vars, etc.
+  log_system_params: true
+
+  # enable/disable tracking console logs (default value is true)
+  capture_terminal_logs: false # set to false to avoid infinite console log loop issue https://github.com/aimhubio/aim/issues/2550

third_party/Matcha-TTS/configs/logger/tensorboard.yaml

@@ -0,0 +1,10 @@
+# https://www.tensorflow.org/tensorboard/
+
+tensorboard:
+  _target_: lightning.pytorch.loggers.tensorboard.TensorBoardLogger
+  save_dir: "${paths.output_dir}/tensorboard/"
+  name: null
+  log_graph: False
+  default_hp_metric: True
+  prefix: ""
+  # version: ""

third_party/Matcha-TTS/configs/logger/wandb.yaml

@@ -0,0 +1,16 @@
+# https://wandb.ai
+
+wandb:
+  _target_: lightning.pytorch.loggers.wandb.WandbLogger
+  # name: "" # name of the run (normally generated by wandb)
+  save_dir: "${paths.output_dir}"
+  offline: False
+  id: null # pass correct id to resume experiment!
+  anonymous: null # enable anonymous logging
+  project: "lightning-hydra-template"
+  log_model: False # upload lightning ckpts
+  prefix: "" # a string to put at the beginning of metric keys
+  # entity: "" # set to name of your wandb team
+  group: ""
+  tags: []
+  job_type: ""

third_party/Matcha-TTS/configs/model/cfm/default.yaml

@@ -0,0 +1,3 @@
+name: CFM
+solver: euler
+sigma_min: 1e-4

third_party/Matcha-TTS/configs/model/decoder/default.yaml

@@ -0,0 +1,7 @@
+channels: [256, 256]
+dropout: 0.05
+attention_head_dim: 64
+n_blocks: 1
+num_mid_blocks: 2
+num_heads: 2
+act_fn: snakebeta

third_party/Matcha-TTS/configs/model/encoder/default.yaml

@@ -0,0 +1,18 @@
+encoder_type: RoPE Encoder
+encoder_params:
+  n_feats: ${model.n_feats}
+  n_channels: 192
+  filter_channels: 768
+  filter_channels_dp: 256
+  n_heads: 2
+  n_layers: 6
+  kernel_size: 3
+  p_dropout: 0.1
+  spk_emb_dim: 64
+  n_spks: 1
+  prenet: true
+
+duration_predictor_params:
+  filter_channels_dp: ${model.encoder.encoder_params.filter_channels_dp}
+  kernel_size: 3
+  p_dropout: ${model.encoder.encoder_params.p_dropout}

third_party/Matcha-TTS/configs/model/matcha.yaml

@@ -0,0 +1,15 @@
+defaults:
+  - _self_
+  - encoder: default.yaml
+  - decoder: default.yaml
+  - cfm: default.yaml
+  - optimizer: adam.yaml
+
+_target_: matcha.models.matcha_tts.MatchaTTS
+n_vocab: 178
+n_spks: ${data.n_spks}
+spk_emb_dim: 64
+n_feats: 80
+data_statistics: ${data.data_statistics}
+out_size: null # Must be divisible by 4
+prior_loss: true

third_party/Matcha-TTS/configs/model/optimizer/adam.yaml

@@ -0,0 +1,4 @@
+_target_: torch.optim.Adam
+_partial_: true
+lr: 1e-4
+weight_decay: 0.0

third_party/Matcha-TTS/configs/paths/default.yaml

@@ -0,0 +1,18 @@
+# path to root directory
+# this requires PROJECT_ROOT environment variable to exist
+# you can replace it with "." if you want the root to be the current working directory
+root_dir: ${oc.env:PROJECT_ROOT}
+
+# path to data directory
+data_dir: ${paths.root_dir}/data/
+
+# path to logging directory
+log_dir: ${paths.root_dir}/logs/
+
+# path to output directory, created dynamically by hydra
+# path generation pattern is specified in `configs/hydra/default.yaml`
+# use it to store all files generated during the run, like ckpts and metrics
+output_dir: ${hydra:runtime.output_dir}
+
+# path to working directory
+work_dir: ${hydra:runtime.cwd}

third_party/Matcha-TTS/configs/train.yaml

@@ -0,0 +1,51 @@
+# @package _global_
+
+# specify here default configuration
+# order of defaults determines the order in which configs override each other
+defaults:
+  - _self_
+  - data: ljspeech
+  - model: matcha
+  - callbacks: default
+  - logger: tensorboard # set logger here or use command line (e.g. `python train.py logger=tensorboard`)
+  - trainer: default
+  - paths: default
+  - extras: default
+  - hydra: default
+
+  # experiment configs allow for version control of specific hyperparameters
+  # e.g. best hyperparameters for given model and datamodule
+  - experiment: null
+
+  # config for hyperparameter optimization
+  - hparams_search: null
+
+  # optional local config for machine/user specific settings
+  # it's optional since it doesn't need to exist and is excluded from version control
+  - optional local: default
+
+  # debugging config (enable through command line, e.g. `python train.py debug=default)
+  - debug: null
+
+# task name, determines output directory path
+task_name: "train"
+
+run_name: ???
+
+# tags to help you identify your experiments
+# you can overwrite this in experiment configs
+# overwrite from command line with `python train.py tags="[first_tag, second_tag]"`
+tags: ["dev"]
+
+# set False to skip model training
+train: True
+
+# evaluate on test set, using best model weights achieved during training
+# lightning chooses best weights based on the metric specified in checkpoint callback
+test: True
+
+# simply provide checkpoint path to resume training
+ckpt_path: null
+
+# seed for random number generators in pytorch, numpy and python.random
+seed: 1234

third_party/Matcha-TTS/configs/trainer/ddp.yaml

@@ -0,0 +1,9 @@
+defaults:
+  - default
+
+strategy: ddp
+
+accelerator: gpu
+devices: [0,1]
+num_nodes: 1
+sync_batchnorm: True

third_party/Matcha-TTS/configs/trainer/ddp_sim.yaml

@@ -0,0 +1,7 @@
+defaults:
+  - default
+
+# simulate DDP on CPU, useful for debugging
+accelerator: cpu
+devices: 2
+strategy: ddp_spawn

third_party/Matcha-TTS/configs/trainer/gpu.yaml

@@ -0,0 +1,5 @@
+defaults:
+  - default
+
+accelerator: gpu
+devices: 1

third_party/Matcha-TTS/configs/trainer/mps.yaml

@@ -0,0 +1,5 @@
+defaults:
+  - default
+
+accelerator: mps
+devices: 1

third_party/Matcha-TTS/matcha/VERSION

@@ -0,0 +1 @@
+0.0.5.1

third_party/Matcha-TTS/matcha/cli.py

@@ -0,0 +1,418 @@
+import argparse
+import datetime as dt
+import os
+import warnings
+from pathlib import Path
+
+import matplotlib.pyplot as plt
+import numpy as np
+import soundfile as sf
+import torch
+
+from matcha.hifigan.config import v1
+from matcha.hifigan.denoiser import Denoiser
+from matcha.hifigan.env import AttrDict
+from matcha.hifigan.models import Generator as HiFiGAN
+from matcha.models.matcha_tts import MatchaTTS
+from matcha.text import sequence_to_text, text_to_sequence
+from matcha.utils.utils import assert_model_downloaded, get_user_data_dir, intersperse
+
+MATCHA_URLS = {
+    "matcha_ljspeech": "https://github.com/shivammehta25/Matcha-TTS-checkpoints/releases/download/v1.0/matcha_ljspeech.ckpt",
+    "matcha_vctk": "https://github.com/shivammehta25/Matcha-TTS-checkpoints/releases/download/v1.0/matcha_vctk.ckpt",
+}
+
+VOCODER_URLS = {
+    "hifigan_T2_v1": "https://github.com/shivammehta25/Matcha-TTS-checkpoints/releases/download/v1.0/generator_v1",  # Old url: https://drive.google.com/file/d/14NENd4equCBLyyCSke114Mv6YR_j_uFs/view?usp=drive_link
+    "hifigan_univ_v1": "https://github.com/shivammehta25/Matcha-TTS-checkpoints/releases/download/v1.0/g_02500000",  # Old url: https://drive.google.com/file/d/1qpgI41wNXFcH-iKq1Y42JlBC9j0je8PW/view?usp=drive_link
+}
+
+MULTISPEAKER_MODEL = {
+    "matcha_vctk": {"vocoder": "hifigan_univ_v1", "speaking_rate": 0.85, "spk": 0, "spk_range": (0, 107)}
+}
+
+SINGLESPEAKER_MODEL = {"matcha_ljspeech": {"vocoder": "hifigan_T2_v1", "speaking_rate": 0.95, "spk": None}}
+
+
+def plot_spectrogram_to_numpy(spectrogram, filename):
+    fig, ax = plt.subplots(figsize=(12, 3))
+    im = ax.imshow(spectrogram, aspect="auto", origin="lower", interpolation="none")
+    plt.colorbar(im, ax=ax)
+    plt.xlabel("Frames")
+    plt.ylabel("Channels")
+    plt.title("Synthesised Mel-Spectrogram")
+    fig.canvas.draw()
+    plt.savefig(filename)
+
+
+def process_text(i: int, text: str, device: torch.device):
+    print(f"[{i}] - Input text: {text}")
+    x = torch.tensor(
+        intersperse(text_to_sequence(text, ["english_cleaners2"]), 0),
+        dtype=torch.long,
+        device=device,
+    )[None]
+    x_lengths = torch.tensor([x.shape[-1]], dtype=torch.long, device=device)
+    x_phones = sequence_to_text(x.squeeze(0).tolist())
+    print(f"[{i}] - Phonetised text: {x_phones[1::2]}")
+
+    return {"x_orig": text, "x": x, "x_lengths": x_lengths, "x_phones": x_phones}
+
+
+def get_texts(args):
+    if args.text:
+        texts = [args.text]
+    else:
+        with open(args.file, encoding="utf-8") as f:
+            texts = f.readlines()
+    return texts
+
+
+def assert_required_models_available(args):
+    save_dir = get_user_data_dir()
+    if not hasattr(args, "checkpoint_path") and args.checkpoint_path is None:
+        model_path = args.checkpoint_path
+    else:
+        model_path = save_dir / f"{args.model}.ckpt"
+        assert_model_downloaded(model_path, MATCHA_URLS[args.model])
+
+    vocoder_path = save_dir / f"{args.vocoder}"
+    assert_model_downloaded(vocoder_path, VOCODER_URLS[args.vocoder])
+    return {"matcha": model_path, "vocoder": vocoder_path}
+
+
+def load_hifigan(checkpoint_path, device):
+    h = AttrDict(v1)
+    hifigan = HiFiGAN(h).to(device)
+    hifigan.load_state_dict(torch.load(checkpoint_path, map_location=device)["generator"])
+    _ = hifigan.eval()
+    hifigan.remove_weight_norm()
+    return hifigan
+
+
+def load_vocoder(vocoder_name, checkpoint_path, device):
+    print(f"[!] Loading {vocoder_name}!")
+    vocoder = None
+    if vocoder_name in ("hifigan_T2_v1", "hifigan_univ_v1"):
+        vocoder = load_hifigan(checkpoint_path, device)
+    else:
+        raise NotImplementedError(
+            f"Vocoder {vocoder_name} not implemented! define a load_<<vocoder_name>> method for it"
+        )
+
+    denoiser = Denoiser(vocoder, mode="zeros")
+    print(f"[+] {vocoder_name} loaded!")
+    return vocoder, denoiser
+
+
+def load_matcha(model_name, checkpoint_path, device):
+    print(f"[!] Loading {model_name}!")
+    model = MatchaTTS.load_from_checkpoint(checkpoint_path, map_location=device)
+    _ = model.eval()
+
+    print(f"[+] {model_name} loaded!")
+    return model
+
+
+def to_waveform(mel, vocoder, denoiser=None):
+    audio = vocoder(mel).clamp(-1, 1)
+    if denoiser is not None:
+        audio = denoiser(audio.squeeze(), strength=0.00025).cpu().squeeze()
+
+    return audio.cpu().squeeze()
+
+
+def save_to_folder(filename: str, output: dict, folder: str):
+    folder = Path(folder)
+    folder.mkdir(exist_ok=True, parents=True)
+    plot_spectrogram_to_numpy(np.array(output["mel"].squeeze().float().cpu()), f"{filename}.png")
+    np.save(folder / f"{filename}", output["mel"].cpu().numpy())
+    sf.write(folder / f"{filename}.wav", output["waveform"], 22050, "PCM_24")
+    return folder.resolve() / f"{filename}.wav"
+
+
+def validate_args(args):
+    assert (
+        args.text or args.file
+    ), "Either text or file must be provided Matcha-T(ea)TTS need sometext to whisk the waveforms."
+    assert args.temperature >= 0, "Sampling temperature cannot be negative"
+    assert args.steps > 0, "Number of ODE steps must be greater than 0"
+
+    if args.checkpoint_path is None:
+        # When using pretrained models
+        if args.model in SINGLESPEAKER_MODEL:
+            args = validate_args_for_single_speaker_model(args)
+
+        if args.model in MULTISPEAKER_MODEL:
+            args = validate_args_for_multispeaker_model(args)
+    else:
+        # When using a custom model
+        if args.vocoder != "hifigan_univ_v1":
+            warn_ = "[-] Using custom model checkpoint! I would suggest passing --vocoder hifigan_univ_v1, unless the custom model is trained on LJ Speech."
+            warnings.warn(warn_, UserWarning)
+        if args.speaking_rate is None:
+            args.speaking_rate = 1.0
+
+    if args.batched:
+        assert args.batch_size > 0, "Batch size must be greater than 0"
+    assert args.speaking_rate > 0, "Speaking rate must be greater than 0"
+
+    return args
+
+
+def validate_args_for_multispeaker_model(args):
+    if args.vocoder is not None:
+        if args.vocoder != MULTISPEAKER_MODEL[args.model]["vocoder"]:
+            warn_ = f"[-] Using {args.model} model! I would suggest passing --vocoder {MULTISPEAKER_MODEL[args.model]['vocoder']}"
+            warnings.warn(warn_, UserWarning)
+    else:
+        args.vocoder = MULTISPEAKER_MODEL[args.model]["vocoder"]
+
+    if args.speaking_rate is None:
+        args.speaking_rate = MULTISPEAKER_MODEL[args.model]["speaking_rate"]
+
+    spk_range = MULTISPEAKER_MODEL[args.model]["spk_range"]
+    if args.spk is not None:
+        assert (
+            args.spk >= spk_range[0] and args.spk <= spk_range[-1]
+        ), f"Speaker ID must be between {spk_range} for this model."
+    else:
+        available_spk_id = MULTISPEAKER_MODEL[args.model]["spk"]
+        warn_ = f"[!] Speaker ID not provided! Using speaker ID {available_spk_id}"
+        warnings.warn(warn_, UserWarning)
+        args.spk = available_spk_id
+
+    return args
+
+
+def validate_args_for_single_speaker_model(args):
+    if args.vocoder is not None:
+        if args.vocoder != SINGLESPEAKER_MODEL[args.model]["vocoder"]:
+            warn_ = f"[-] Using {args.model} model! I would suggest passing --vocoder {SINGLESPEAKER_MODEL[args.model]['vocoder']}"
+            warnings.warn(warn_, UserWarning)
+    else:
+        args.vocoder = SINGLESPEAKER_MODEL[args.model]["vocoder"]
+
+    if args.speaking_rate is None:
+        args.speaking_rate = SINGLESPEAKER_MODEL[args.model]["speaking_rate"]
+
+    if args.spk != SINGLESPEAKER_MODEL[args.model]["spk"]:
+        warn_ = f"[-] Ignoring speaker id {args.spk} for {args.model}"
+        warnings.warn(warn_, UserWarning)
+        args.spk = SINGLESPEAKER_MODEL[args.model]["spk"]
+
+    return args
+
+
+@torch.inference_mode()
+def cli():
+    parser = argparse.ArgumentParser(
+        description=" 🍵 Matcha-TTS: A fast TTS architecture with conditional flow matching"
+    )
+    parser.add_argument(
+        "--model",
+        type=str,
+        default="matcha_ljspeech",
+        help="Model to use",
+        choices=MATCHA_URLS.keys(),
+    )
+
+    parser.add_argument(
+        "--checkpoint_path",
+        type=str,
+        default=None,
+        help="Path to the custom model checkpoint",
+    )
+
+    parser.add_argument(
+        "--vocoder",
+        type=str,
+        default=None,
+        help="Vocoder to use (default: will use the one suggested with the pretrained model))",
+        choices=VOCODER_URLS.keys(),
+    )
+    parser.add_argument("--text", type=str, default=None, help="Text to synthesize")
+    parser.add_argument("--file", type=str, default=None, help="Text file to synthesize")
+    parser.add_argument("--spk", type=int, default=None, help="Speaker ID")
+    parser.add_argument(
+        "--temperature",
+        type=float,
+        default=0.667,
+        help="Variance of the x0 noise (default: 0.667)",
+    )
+    parser.add_argument(
+        "--speaking_rate",
+        type=float,
+        default=None,
+        help="change the speaking rate, a higher value means slower speaking rate (default: 1.0)",
+    )
+    parser.add_argument("--steps", type=int, default=10, help="Number of ODE steps  (default: 10)")
+    parser.add_argument("--cpu", action="store_true", help="Use CPU for inference (default: use GPU if available)")
+    parser.add_argument(
+        "--denoiser_strength",
+        type=float,
+        default=0.00025,
+        help="Strength of the vocoder bias denoiser (default: 0.00025)",
+    )
+    parser.add_argument(
+        "--output_folder",
+        type=str,
+        default=os.getcwd(),
+        help="Output folder to save results (default: current dir)",
+    )
+    parser.add_argument("--batched", action="store_true", help="Batched inference (default: False)")
+    parser.add_argument(
+        "--batch_size", type=int, default=32, help="Batch size only useful when --batched (default: 32)"
+    )
+
+    args = parser.parse_args()
+
+    args = validate_args(args)
+    device = get_device(args)
+    print_config(args)
+    paths = assert_required_models_available(args)
+
+    if args.checkpoint_path is not None:
+        print(f"[🍵] Loading custom model from {args.checkpoint_path}")
+        paths["matcha"] = args.checkpoint_path
+        args.model = "custom_model"
+
+    model = load_matcha(args.model, paths["matcha"], device)
+    vocoder, denoiser = load_vocoder(args.vocoder, paths["vocoder"], device)
+
+    texts = get_texts(args)
+
+    spk = torch.tensor([args.spk], device=device, dtype=torch.long) if args.spk is not None else None
+    if len(texts) == 1 or not args.batched:
+        unbatched_synthesis(args, device, model, vocoder, denoiser, texts, spk)
+    else:
+        batched_synthesis(args, device, model, vocoder, denoiser, texts, spk)
+
+
+class BatchedSynthesisDataset(torch.utils.data.Dataset):
+    def __init__(self, processed_texts):
+        self.processed_texts = processed_texts
+
+    def __len__(self):
+        return len(self.processed_texts)
+
+    def __getitem__(self, idx):
+        return self.processed_texts[idx]
+
+
+def batched_collate_fn(batch):
+    x = []
+    x_lengths = []
+
+    for b in batch:
+        x.append(b["x"].squeeze(0))
+        x_lengths.append(b["x_lengths"])
+
+    x = torch.nn.utils.rnn.pad_sequence(x, batch_first=True)
+    x_lengths = torch.concat(x_lengths, dim=0)
+    return {"x": x, "x_lengths": x_lengths}
+
+
+def batched_synthesis(args, device, model, vocoder, denoiser, texts, spk):
+    total_rtf = []
+    total_rtf_w = []
+    processed_text = [process_text(i, text, "cpu") for i, text in enumerate(texts)]
+    dataloader = torch.utils.data.DataLoader(
+        BatchedSynthesisDataset(processed_text),
+        batch_size=args.batch_size,
+        collate_fn=batched_collate_fn,
+        num_workers=8,
+    )
+    for i, batch in enumerate(dataloader):
+        i = i + 1
+        start_t = dt.datetime.now()
+        output = model.synthesise(
+            batch["x"].to(device),
+            batch["x_lengths"].to(device),
+            n_timesteps=args.steps,
+            temperature=args.temperature,
+            spks=spk,
+            length_scale=args.speaking_rate,
+        )
+
+        output["waveform"] = to_waveform(output["mel"], vocoder, denoiser)
+        t = (dt.datetime.now() - start_t).total_seconds()
+        rtf_w = t * 22050 / (output["waveform"].shape[-1])
+        print(f"[🍵-Batch: {i}] Matcha-TTS RTF: {output['rtf']:.4f}")
+        print(f"[🍵-Batch: {i}] Matcha-TTS + VOCODER RTF: {rtf_w:.4f}")
+        total_rtf.append(output["rtf"])
+        total_rtf_w.append(rtf_w)
+        for j in range(output["mel"].shape[0]):
+            base_name = f"utterance_{j:03d}_speaker_{args.spk:03d}" if args.spk is not None else f"utterance_{j:03d}"
+            length = output["mel_lengths"][j]
+            new_dict = {"mel": output["mel"][j][:, :length], "waveform": output["waveform"][j][: length * 256]}
+            location = save_to_folder(base_name, new_dict, args.output_folder)
+            print(f"[🍵-{j}] Waveform saved: {location}")
+
+    print("".join(["="] * 100))
+    print(f"[🍵] Average Matcha-TTS RTF: {np.mean(total_rtf):.4f} ± {np.std(total_rtf)}")
+    print(f"[🍵] Average Matcha-TTS + VOCODER RTF: {np.mean(total_rtf_w):.4f} ± {np.std(total_rtf_w)}")
+    print("[🍵] Enjoy the freshly whisked 🍵 Matcha-TTS!")
+
+
+def unbatched_synthesis(args, device, model, vocoder, denoiser, texts, spk):
+    total_rtf = []
+    total_rtf_w = []
+    for i, text in enumerate(texts):
+        i = i + 1
+        base_name = f"utterance_{i:03d}_speaker_{args.spk:03d}" if args.spk is not None else f"utterance_{i:03d}"
+
+        print("".join(["="] * 100))
+        text = text.strip()
+        text_processed = process_text(i, text, device)
+
+        print(f"[🍵] Whisking Matcha-T(ea)TS for: {i}")
+        start_t = dt.datetime.now()
+        output = model.synthesise(
+            text_processed["x"],
+            text_processed["x_lengths"],
+            n_timesteps=args.steps,
+            temperature=args.temperature,
+            spks=spk,
+            length_scale=args.speaking_rate,
+        )
+        output["waveform"] = to_waveform(output["mel"], vocoder, denoiser)
+        # RTF with HiFiGAN
+        t = (dt.datetime.now() - start_t).total_seconds()
+        rtf_w = t * 22050 / (output["waveform"].shape[-1])
+        print(f"[🍵-{i}] Matcha-TTS RTF: {output['rtf']:.4f}")
+        print(f"[🍵-{i}] Matcha-TTS + VOCODER RTF: {rtf_w:.4f}")
+        total_rtf.append(output["rtf"])
+        total_rtf_w.append(rtf_w)
+
+        location = save_to_folder(base_name, output, args.output_folder)
+        print(f"[+] Waveform saved: {location}")
+
+    print("".join(["="] * 100))
+    print(f"[🍵] Average Matcha-TTS RTF: {np.mean(total_rtf):.4f} ± {np.std(total_rtf)}")
+    print(f"[🍵] Average Matcha-TTS + VOCODER RTF: {np.mean(total_rtf_w):.4f} ± {np.std(total_rtf_w)}")
+    print("[🍵] Enjoy the freshly whisked 🍵 Matcha-TTS!")
+
+
+def print_config(args):
+    print("[!] Configurations: ")
+    print(f"\t- Model: {args.model}")
+    print(f"\t- Vocoder: {args.vocoder}")
+    print(f"\t- Temperature: {args.temperature}")
+    print(f"\t- Speaking rate: {args.speaking_rate}")
+    print(f"\t- Number of ODE steps: {args.steps}")
+    print(f"\t- Speaker: {args.spk}")
+
+
+def get_device(args):
+    if torch.cuda.is_available() and not args.cpu:
+        print("[+] GPU Available! Using GPU")
+        device = torch.device("cuda")
+    else:
+        print("[-] GPU not available or forced CPU run! Using CPU")
+        device = torch.device("cpu")
+    return device
+
+
+if __name__ == "__main__":
+    cli()

third_party/Matcha-TTS/matcha/hifigan/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2020 Jungil Kong
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

third_party/Matcha-TTS/matcha/hifigan/meldataset.py

@@ -0,0 +1,217 @@
+""" from https://github.com/jik876/hifi-gan """
+
+import math
+import os
+import random
+
+import numpy as np
+import torch
+import torch.utils.data
+from librosa.filters import mel as librosa_mel_fn
+from librosa.util import normalize
+from scipy.io.wavfile import read
+
+MAX_WAV_VALUE = 32768.0
+
+
+def load_wav(full_path):
+    sampling_rate, data = read(full_path)
+    return data, sampling_rate
+
+
+def dynamic_range_compression(x, C=1, clip_val=1e-5):
+    return np.log(np.clip(x, a_min=clip_val, a_max=None) * C)
+
+
+def dynamic_range_decompression(x, C=1):
+    return np.exp(x) / C
+
+
+def dynamic_range_compression_torch(x, C=1, clip_val=1e-5):
+    return torch.log(torch.clamp(x, min=clip_val) * C)
+
+
+def dynamic_range_decompression_torch(x, C=1):
+    return torch.exp(x) / C
+
+
+def spectral_normalize_torch(magnitudes):
+    output = dynamic_range_compression_torch(magnitudes)
+    return output
+
+
+def spectral_de_normalize_torch(magnitudes):
+    output = dynamic_range_decompression_torch(magnitudes)
+    return output
+
+
+mel_basis = {}
+hann_window = {}
+
+
+def mel_spectrogram(y, n_fft, num_mels, sampling_rate, hop_size, win_size, fmin, fmax, center=False):
+    if torch.min(y) < -1.0:
+        print("min value is ", torch.min(y))
+    if torch.max(y) > 1.0:
+        print("max value is ", torch.max(y))
+
+    global mel_basis, hann_window  # pylint: disable=global-statement
+    if fmax not in mel_basis:
+        mel = librosa_mel_fn(sampling_rate, n_fft, num_mels, fmin, fmax)
+        mel_basis[str(fmax) + "_" + str(y.device)] = torch.from_numpy(mel).float().to(y.device)
+        hann_window[str(y.device)] = torch.hann_window(win_size).to(y.device)
+
+    y = torch.nn.functional.pad(
+        y.unsqueeze(1), (int((n_fft - hop_size) / 2), int((n_fft - hop_size) / 2)), mode="reflect"
+    )
+    y = y.squeeze(1)
+
+    spec = torch.view_as_real(
+        torch.stft(
+            y,
+            n_fft,
+            hop_length=hop_size,
+            win_length=win_size,
+            window=hann_window[str(y.device)],
+            center=center,
+            pad_mode="reflect",
+            normalized=False,
+            onesided=True,
+            return_complex=True,
+        )
+    )
+
+    spec = torch.sqrt(spec.pow(2).sum(-1) + (1e-9))
+
+    spec = torch.matmul(mel_basis[str(fmax) + "_" + str(y.device)], spec)
+    spec = spectral_normalize_torch(spec)
+
+    return spec
+
+
+def get_dataset_filelist(a):
+    with open(a.input_training_file, encoding="utf-8") as fi:
+        training_files = [
+            os.path.join(a.input_wavs_dir, x.split("|")[0] + ".wav") for x in fi.read().split("\n") if len(x) > 0
+        ]
+
+    with open(a.input_validation_file, encoding="utf-8") as fi:
+        validation_files = [
+            os.path.join(a.input_wavs_dir, x.split("|")[0] + ".wav") for x in fi.read().split("\n") if len(x) > 0
+        ]
+    return training_files, validation_files
+
+
+class MelDataset(torch.utils.data.Dataset):
+    def __init__(
+        self,
+        training_files,
+        segment_size,
+        n_fft,
+        num_mels,
+        hop_size,
+        win_size,
+        sampling_rate,
+        fmin,
+        fmax,
+        split=True,
+        shuffle=True,
+        n_cache_reuse=1,
+        device=None,
+        fmax_loss=None,
+        fine_tuning=False,
+        base_mels_path=None,
+    ):
+        self.audio_files = training_files
+        random.seed(1234)
+        if shuffle:
+            random.shuffle(self.audio_files)
+        self.segment_size = segment_size
+        self.sampling_rate = sampling_rate
+        self.split = split
+        self.n_fft = n_fft
+        self.num_mels = num_mels
+        self.hop_size = hop_size
+        self.win_size = win_size
+        self.fmin = fmin
+        self.fmax = fmax
+        self.fmax_loss = fmax_loss
+        self.cached_wav = None
+        self.n_cache_reuse = n_cache_reuse
+        self._cache_ref_count = 0
+        self.device = device
+        self.fine_tuning = fine_tuning
+        self.base_mels_path = base_mels_path
+
+    def __getitem__(self, index):
+        filename = self.audio_files[index]
+        if self._cache_ref_count == 0:
+            audio, sampling_rate = load_wav(filename)
+            audio = audio / MAX_WAV_VALUE
+            if not self.fine_tuning:
+                audio = normalize(audio) * 0.95
+            self.cached_wav = audio
+            if sampling_rate != self.sampling_rate:
+                raise ValueError(f"{sampling_rate} SR doesn't match target {self.sampling_rate} SR")
+            self._cache_ref_count = self.n_cache_reuse
+        else:
+            audio = self.cached_wav
+            self._cache_ref_count -= 1
+
+        audio = torch.FloatTensor(audio)
+        audio = audio.unsqueeze(0)
+
+        if not self.fine_tuning:
+            if self.split:
+                if audio.size(1) >= self.segment_size:
+                    max_audio_start = audio.size(1) - self.segment_size
+                    audio_start = random.randint(0, max_audio_start)
+                    audio = audio[:, audio_start : audio_start + self.segment_size]
+                else:
+                    audio = torch.nn.functional.pad(audio, (0, self.segment_size - audio.size(1)), "constant")
+
+            mel = mel_spectrogram(
+                audio,
+                self.n_fft,
+                self.num_mels,
+                self.sampling_rate,
+                self.hop_size,
+                self.win_size,
+                self.fmin,
+                self.fmax,
+                center=False,
+            )
+        else:
+            mel = np.load(os.path.join(self.base_mels_path, os.path.splitext(os.path.split(filename)[-1])[0] + ".npy"))
+            mel = torch.from_numpy(mel)
+
+            if len(mel.shape) < 3:
+                mel = mel.unsqueeze(0)
+
+            if self.split:
+                frames_per_seg = math.ceil(self.segment_size / self.hop_size)
+
+                if audio.size(1) >= self.segment_size:
+                    mel_start = random.randint(0, mel.size(2) - frames_per_seg - 1)
+                    mel = mel[:, :, mel_start : mel_start + frames_per_seg]
+                    audio = audio[:, mel_start * self.hop_size : (mel_start + frames_per_seg) * self.hop_size]
+                else:
+                    mel = torch.nn.functional.pad(mel, (0, frames_per_seg - mel.size(2)), "constant")
+                    audio = torch.nn.functional.pad(audio, (0, self.segment_size - audio.size(1)), "constant")
+
+        mel_loss = mel_spectrogram(
+            audio,
+            self.n_fft,
+            self.num_mels,
+            self.sampling_rate,
+            self.hop_size,
+            self.win_size,
+            self.fmin,
+            self.fmax_loss,
+            center=False,
+        )
+
+        return (mel.squeeze(), audio.squeeze(0), filename, mel_loss.squeeze())
+
+    def __len__(self):
+        return len(self.audio_files)

third_party/Matcha-TTS/matcha/onnx/infer.py

@@ -0,0 +1,168 @@
+import argparse
+import os
+import warnings
+from pathlib import Path
+from time import perf_counter
+
+import numpy as np
+import onnxruntime as ort
+import soundfile as sf
+import torch
+
+from matcha.cli import plot_spectrogram_to_numpy, process_text
+
+
+def validate_args(args):
+    assert (
+        args.text or args.file
+    ), "Either text or file must be provided Matcha-T(ea)TTS need sometext to whisk the waveforms."
+    assert args.temperature >= 0, "Sampling temperature cannot be negative"
+    assert args.speaking_rate >= 0, "Speaking rate must be greater than 0"
+    return args
+
+
+def write_wavs(model, inputs, output_dir, external_vocoder=None):
+    if external_vocoder is None:
+        print("The provided model has the vocoder embedded in the graph.\nGenerating waveform directly")
+        t0 = perf_counter()
+        wavs, wav_lengths = model.run(None, inputs)
+        infer_secs = perf_counter() - t0
+        mel_infer_secs = vocoder_infer_secs = None
+    else:
+        print("[🍵] Generating mel using Matcha")
+        mel_t0 = perf_counter()
+        mels, mel_lengths = model.run(None, inputs)
+        mel_infer_secs = perf_counter() - mel_t0
+        print("Generating waveform from mel using external vocoder")
+        vocoder_inputs = {external_vocoder.get_inputs()[0].name: mels}
+        vocoder_t0 = perf_counter()
+        wavs = external_vocoder.run(None, vocoder_inputs)[0]
+        vocoder_infer_secs = perf_counter() - vocoder_t0
+        wavs = wavs.squeeze(1)
+        wav_lengths = mel_lengths * 256
+        infer_secs = mel_infer_secs + vocoder_infer_secs
+
+    output_dir = Path(output_dir)
+    output_dir.mkdir(parents=True, exist_ok=True)
+    for i, (wav, wav_length) in enumerate(zip(wavs, wav_lengths)):
+        output_filename = output_dir.joinpath(f"output_{i + 1}.wav")
+        audio = wav[:wav_length]
+        print(f"Writing audio to {output_filename}")
+        sf.write(output_filename, audio, 22050, "PCM_24")
+
+    wav_secs = wav_lengths.sum() / 22050
+    print(f"Inference seconds: {infer_secs}")
+    print(f"Generated wav seconds: {wav_secs}")
+    rtf = infer_secs / wav_secs
+    if mel_infer_secs is not None:
+        mel_rtf = mel_infer_secs / wav_secs
+        print(f"Matcha RTF: {mel_rtf}")
+    if vocoder_infer_secs is not None:
+        vocoder_rtf = vocoder_infer_secs / wav_secs
+        print(f"Vocoder RTF: {vocoder_rtf}")
+    print(f"Overall RTF: {rtf}")
+
+
+def write_mels(model, inputs, output_dir):
+    t0 = perf_counter()
+    mels, mel_lengths = model.run(None, inputs)
+    infer_secs = perf_counter() - t0
+
+    output_dir = Path(output_dir)
+    output_dir.mkdir(parents=True, exist_ok=True)
+    for i, mel in enumerate(mels):
+        output_stem = output_dir.joinpath(f"output_{i + 1}")
+        plot_spectrogram_to_numpy(mel.squeeze(), output_stem.with_suffix(".png"))
+        np.save(output_stem.with_suffix(".numpy"), mel)
+
+    wav_secs = (mel_lengths * 256).sum() / 22050
+    print(f"Inference seconds: {infer_secs}")
+    print(f"Generated wav seconds: {wav_secs}")
+    rtf = infer_secs / wav_secs
+    print(f"RTF: {rtf}")
+
+
+def main():
+    parser = argparse.ArgumentParser(
+        description=" 🍵 Matcha-TTS: A fast TTS architecture with conditional flow matching"
+    )
+    parser.add_argument(
+        "model",
+        type=str,
+        help="ONNX model to use",
+    )
+    parser.add_argument("--vocoder", type=str, default=None, help="Vocoder to use (defaults to None)")
+    parser.add_argument("--text", type=str, default=None, help="Text to synthesize")
+    parser.add_argument("--file", type=str, default=None, help="Text file to synthesize")
+    parser.add_argument("--spk", type=int, default=None, help="Speaker ID")
+    parser.add_argument(
+        "--temperature",
+        type=float,
+        default=0.667,
+        help="Variance of the x0 noise (default: 0.667)",
+    )
+    parser.add_argument(
+        "--speaking-rate",
+        type=float,
+        default=1.0,
+        help="change the speaking rate, a higher value means slower speaking rate (default: 1.0)",
+    )
+    parser.add_argument("--gpu", action="store_true", help="Use CPU for inference (default: use GPU if available)")
+    parser.add_argument(
+        "--output-dir",
+        type=str,
+        default=os.getcwd(),
+        help="Output folder to save results (default: current dir)",
+    )
+
+    args = parser.parse_args()
+    args = validate_args(args)
+
+    if args.gpu:
+        providers = ["GPUExecutionProvider"]
+    else:
+        providers = ["CPUExecutionProvider"]
+    model = ort.InferenceSession(args.model, providers=providers)
+
+    model_inputs = model.get_inputs()
+    model_outputs = list(model.get_outputs())
+
+    if args.text:
+        text_lines = args.text.splitlines()
+    else:
+        with open(args.file, encoding="utf-8") as file:
+            text_lines = file.read().splitlines()
+
+    processed_lines = [process_text(0, line, "cpu") for line in text_lines]
+    x = [line["x"].squeeze() for line in processed_lines]
+    # Pad
+    x = torch.nn.utils.rnn.pad_sequence(x, batch_first=True)
+    x = x.detach().cpu().numpy()
+    x_lengths = np.array([line["x_lengths"].item() for line in processed_lines], dtype=np.int64)
+    inputs = {
+        "x": x,
+        "x_lengths": x_lengths,
+        "scales": np.array([args.temperature, args.speaking_rate], dtype=np.float32),
+    }
+    is_multi_speaker = len(model_inputs) == 4
+    if is_multi_speaker:
+        if args.spk is None:
+            args.spk = 0
+            warn = "[!] Speaker ID not provided! Using speaker ID 0"
+            warnings.warn(warn, UserWarning)
+        inputs["spks"] = np.repeat(args.spk, x.shape[0]).astype(np.int64)
+
+    has_vocoder_embedded = model_outputs[0].name == "wav"
+    if has_vocoder_embedded:
+        write_wavs(model, inputs, args.output_dir)
+    elif args.vocoder:
+        external_vocoder = ort.InferenceSession(args.vocoder, providers=providers)
+        write_wavs(model, inputs, args.output_dir, external_vocoder=external_vocoder)
+    else:
+        warn = "[!] A vocoder is not embedded in the graph nor an external vocoder is provided. The mel output will be written as numpy arrays to `*.npy` files in the output directory"
+        warnings.warn(warn, UserWarning)
+        write_mels(model, inputs, args.output_dir)
+
+
+if __name__ == "__main__":
+    main()

third_party/Matcha-TTS/matcha/text/__init__.py

@@ -0,0 +1,53 @@
+""" from https://github.com/keithito/tacotron """
+from matcha.text import cleaners
+from matcha.text.symbols import symbols
+
+# Mappings from symbol to numeric ID and vice versa:
+_symbol_to_id = {s: i for i, s in enumerate(symbols)}
+_id_to_symbol = {i: s for i, s in enumerate(symbols)}  # pylint: disable=unnecessary-comprehension
+
+
+def text_to_sequence(text, cleaner_names):
+    """Converts a string of text to a sequence of IDs corresponding to the symbols in the text.
+    Args:
+      text: string to convert to a sequence
+      cleaner_names: names of the cleaner functions to run the text through
+    Returns:
+      List of integers corresponding to the symbols in the text
+    """
+    sequence = []
+
+    clean_text = _clean_text(text, cleaner_names)
+    for symbol in clean_text:
+        symbol_id = _symbol_to_id[symbol]
+        sequence += [symbol_id]
+    return sequence
+
+
+def cleaned_text_to_sequence(cleaned_text):
+    """Converts a string of text to a sequence of IDs corresponding to the symbols in the text.
+    Args:
+      text: string to convert to a sequence
+    Returns:
+      List of integers corresponding to the symbols in the text
+    """
+    sequence = [_symbol_to_id[symbol] for symbol in cleaned_text]
+    return sequence
+
+
+def sequence_to_text(sequence):
+    """Converts a sequence of IDs back to a string"""
+    result = ""
+    for symbol_id in sequence:
+        s = _id_to_symbol[symbol_id]
+        result += s
+    return result
+
+
+def _clean_text(text, cleaner_names):
+    for name in cleaner_names:
+        cleaner = getattr(cleaners, name)
+        if not cleaner:
+            raise Exception("Unknown cleaner: %s" % name)
+        text = cleaner(text)
+    return text