Create app.py

app.py

@@ -0,0 +1,183 @@
+import os
+import torch
+import librosa
+import gradio as gr
+from scipy.io.wavfile import write
+from huggingface_hub import hf_hub_download, snapshot_download
+import utils
+from models import SynthesizerTrn
+from mel_processing import mel_spectrogram_torch
+from speaker_encoder.voice_encoder import SpeakerEncoder
+import logging
+from transformers import Wav2Vec2FeatureExtractor, HubertModel
+
+# 设置日志级别
+logging.getLogger('numba').setLevel(logging.WARNING)
+
+# 模型配置
+MODEL_CONFIG = {
+    "freevc": {
+        "repo_id": "guetLzy/Chinese-FreeVC-Model",  # FreeVC模型仓库
+        "files": ["G_17000.pth", "G_35000.pth"]  # 需要下载的模型文件
+    },
+    "hubert": {
+        "repo_id": "guetLzy/chinese-hubert-large-fariseq-ckpt",  # 中文HuBERT官方仓库
+    }
+}
+
+# 设备设置
+device = 'cuda' if torch.cuda.is_available() else 'cpu'
+
+# 可用的模型选项
+MODEL_OPTIONS = {
+    "Model_17000": "model/G_17000.pth",
+    "Model_35000": "model/G_35000.pth",  # 新增 G_35000.pth 选项
+}
+
+def download_models():
+    """下载所有需要的模型文件"""
+    os.makedirs("model", exist_ok=True)
+    os.makedirs("hubert/chinese-hubert-large-fairseq-ckpt", exist_ok=True)
+
+    # 下载FreeVC模型
+    freevc_paths = {}
+    for model_name, model_path in MODEL_OPTIONS.items():
+        path = hf_hub_download(
+            repo_id=MODEL_CONFIG["freevc"]["repo_id"],
+            filename=os.path.basename(model_path),
+            local_dir="model",
+            resume_download=True
+        )
+        freevc_paths[model_name] = path
+
+    # 下载整个HuBERT仓库
+    hubert_dir = "hubert/chinese-hubert-large-fairseq-ckpt"
+    snapshot_download(
+        repo_id=MODEL_CONFIG["hubert"]["repo_id"],
+        local_dir=hubert_dir,
+        repo_type="model",
+        resume_download=True
+    )
+    hubert_paths = {"snapshot": hubert_dir}  # 返回整个目录路径
+
+    return {
+        "freevc": freevc_paths,
+        "hubert": hubert_paths
+    }
+
+def load_hubert(hubert_dir):
+    """加载HuBERT模型（使用fairseq格式的检查点）"""
+    print("正在加载 HuBERT 模型...")
+
+    # 加载标准HuBERT模型
+    model = HubertModel.from_pretrained(hubert_dir)
+    feature_extractor = Wav2Vec2FeatureExtractor.from_pretrained(hubert_dir)
+
+    return model.to(device).float().eval(), feature_extractor
+
+def load_freevc(model_path):
+    """加载FreeVC模型（使用本地配置文件）"""
+    print(f"正在从 {model_path} 加载 FreeVC 模型...")
+    hps = utils.get_hparams_from_file("configs/freevc.json")  # 本地配置文件
+
+    net_g = SynthesizerTrn(
+        hps.data.filter_length // 2 + 1,
+        hps.train.segment_size // hps.data.hop_length,
+        **hps.model
+    ).to(device)
+
+    utils.load_checkpoint(model_path, net_g, None, True)
+    net_g.eval()
+
+    # 加载本地说话人编码器
+    smodel = SpeakerEncoder("speaker_encoder/ckpt/pretrained_bak_5805000.pt") if hps.model.use_spk else None
+    return net_g, smodel, hps
+
+# 预加载模型
+print("正在下载模型...")
+model_paths = download_models()
+print(model_paths)
+print("正在初始化 HuBERT...")
+hubert_dir = "hubert/chinese-hubert-large-fairseq-ckpt"
+hubert_model, hubert_feature_extractor = load_hubert(hubert_dir)
+
+def voice_conversion(src_audio, tgt_audio, output_name, model_selection):
+    """执行语音转换"""
+    try:
+        # 加载选中的FreeVC模型
+        freevc_model, speaker_model, hps = load_freevc(MODEL_OPTIONS[model_selection])
+
+        with torch.no_grad():
+            # 处理目标音频
+            wav_tgt, _ = librosa.load(tgt_audio, sr=hps.data.sampling_rate)
+            wav_tgt, _ = librosa.effects.trim(wav_tgt, top_db=20)
+            
+            if hps.model.use_spk:
+                g_tgt = speaker_model.embed_utterance(wav_tgt)
+                g_tgt = torch.from_numpy(g_tgt).unsqueeze(0).to(device)
+            else:
+                wav_tgt = torch.from_numpy(wav_tgt).unsqueeze(0).to(device)
+                mel_tgt = mel_spectrogram_torch(
+                    wav_tgt, 
+                    hps.data.filter_length,
+                    hps.data.n_mel_channels,
+                    hps.data.sampling_rate,
+                    hps.data.hop_length,
+                    hps.data.win_length,
+                    hps.data.mel_fmin,
+                    hps.data.mel_fmax
+                )
+
+            # 处理源音频（HuBERT需要16kHz）
+            wav_src, _ = librosa.load(src_audio, sr=16000)
+            inputs = hubert_feature_extractor(
+                wav_src, 
+                return_tensors="pt",
+                sampling_rate=16000
+            ).input_values.to(device)
+            
+            c = hubert_model(inputs.float()).last_hidden_state.transpose(1, 2)
+
+            # 执行转换
+            audio = freevc_model.infer(c, g=g_tgt) if hps.model.use_spk else freevc_model.infer(c, mel=mel_tgt)
+            
+            # 保存结果
+            os.makedirs("output", exist_ok=True)
+            output_path = f"output/{output_name}.wav"
+            write(output_path, hps.data.sampling_rate, audio[0][0].data.cpu().float().numpy())
+            
+            return output_path
+        
+    except Exception as e:
+        print(f"转换错误: {str(e)}")
+        return None
+
+# Gradio界面
+with gr.Blocks(title="Chinese-FreeVC 语音转换") as app:
+    gr.Markdown("## Chinese-FreeVC 语音转换系统")
+
+    with gr.Row():
+        with gr.Column():
+            src_input = gr.Audio(label="源语音", type="filepath")
+            tgt_input = gr.Audio(label="目标音色", type="filepath")
+            model_dropdown = gr.Dropdown(
+                choices=list(MODEL_OPTIONS.keys()),
+                label="选择模型",
+                value="Model_17000"
+            )
+            output_name = gr.Textbox(label="输出文件名", value="converted")
+            convert_btn = gr.Button("开始转换", variant="primary")
+        
+        with gr.Column():
+            output_audio = gr.Audio(label="转换结果", interactive=False)
+            status = gr.Textbox(label="状态")
+
+    convert_btn.click(
+        fn=voice_conversion,
+        inputs=[src_input, tgt_input, output_name, model_dropdown],
+        outputs=[output_audio],
+        api_name="convert"
+    )
+
+if __name__ == "__main__":
+    app.launch(server_name="0.0.0.0", server_port=7860)