import spaces import gradio as gr import torch from transformers.models.speecht5.number_normalizer import EnglishNumberNormalizer from string import punctuation import re from parler_tts import ParlerTTSForConditionalGeneration from transformers import AutoTokenizer, AutoFeatureExtractor, set_seed device = "cuda:0" if torch.cuda.is_available() else "cpu" repo_id = "ylacombe/p-m-e" model = ParlerTTSForConditionalGeneration.from_pretrained(repo_id).to(device) text_tokenizer = AutoTokenizer.from_pretrained(repo_id) description_tokenizer = AutoTokenizer.from_pretrained("google/flan-t5-large") feature_extractor = AutoFeatureExtractor.from_pretrained(repo_id) SAMPLE_RATE = feature_extractor.sampling_rate SEED = 42 default_text = "La voix humaine est un instrument de musique au-dessus de tous les autres." default_description = "a woman with a slightly low- pitched voice speaks slowly in a clear and close- sounding environment, but her delivery is quite monotone." examples = [ # French [ "La voix humaine est un instrument de musique au-dessus de tous les autres.", "a woman with a slightly low- pitched voice speaks slowly in a clear and close- sounding environment, but her delivery is quite monotone.", None, ], # Spanish [ "La voz es el reflejo del alma en el espejo del tiempo.", "a man with a moderate pitch voice speaks slowly with a slightly animated delivery in a very close- sounding environment with minimal background noise.", None, ], # Italian [ "La voce umana è la più bella musica che esista al mondo.", "a man with a moderate pitch speaks slowly in a very noisy environment that sounds very distant, delivering his words in a monotone manner.", None, ], # Portuguese [ "A voz é o espelho da alma e o som do coração.", "a man speaks slowly in a distant- sounding environment with a clean audio quality, delivering his message in a monotone voice at a moderate pitch. ", None, ], # Polish [ "Głos ludzki jest najpiękniejszym instrumentem świata.", "a man with a moderate pitch speaks in a monotone manner at a slightly slow pace, but the recording is quite noisy and sounds very distant.", None, ], # German [ "Die menschliche Stimme ist das schönste Instrument der Welt.", "a man with a moderate pitch speaks slowly in a noisy environment with a flat tone of voice, creating a slightly close- sounding effect.", None, ], # Dutch [ "De menselijke stem is het mooiste instrument dat er bestaat.", "a man with a moderate pitch speaks slightly slowly with an expressive and animated delivery in a very close- sounding environment with a bit of background noise.", None, ], # English [ "The human voice is nature's most perfect instrument.", "Aa woman with a slightly low- pitched voice speaks slowly in a very distant- sounding environment with a clean audio quality, delivering her message in a very monotone manner.", None, ], ] number_normalizer = EnglishNumberNormalizer() def preprocess(text): text = number_normalizer(text).strip() text = text.replace("-", " ") if text[-1] not in punctuation: text = f"{text}." abbreviations_pattern = r'\b[A-Z][A-Z\.]+\b' def separate_abb(chunk): chunk = chunk.replace(".","") print(chunk) return " ".join(chunk) abbreviations = re.findall(abbreviations_pattern, text) for abv in abbreviations: if abv in text: text = text.replace(abv, separate_abb(abv)) return text @spaces.GPU def gen_tts(text, description): inputs = description_tokenizer(description.strip(), return_tensors="pt").to(device) prompt = text_tokenizer(preprocess(text), return_tensors="pt").to(device) set_seed(SEED) generation = model.generate( input_ids=inputs.input_ids, prompt_input_ids=prompt.input_ids, attention_mask=inputs.attention_mask, prompt_attention_mask=prompt.attention_mask, do_sample=True, temperature=1.0 ) audio_arr = generation.cpu().numpy().squeeze() return SAMPLE_RATE, audio_arr css = """ #share-btn-container { display: flex; padding-left: 0.5rem !important; padding-right: 0.5rem !important; background-color: #000000; justify-content: center; align-items: center; border-radius: 9999px !important; width: 13rem; margin-top: 10px; margin-left: auto; flex: unset !important; } #share-btn { all: initial; color: #ffffff; font-weight: 600; cursor: pointer; font-family: 'IBM Plex Sans', sans-serif; margin-left: 0.5rem !important; padding-top: 0.25rem !important; padding-bottom: 0.25rem !important; right:0; } #share-btn * { all: unset !important; } #share-btn-container div:nth-child(-n+2){ width: auto !important; min-height: 0px !important; } #share-btn-container .wrap { display: none !important; } """ with gr.Blocks(css=css) as block: gr.HTML( """

Multi Parler-TTS 🗣️

""" ) gr.HTML( f"""

Parler-TTS is a training and inference library for high-fidelity text-to-speech (TTS) models.

This multilingual model supports French, Spanish, Italian, Portuguese, Polish, German, Dutch, and English. It generates high-quality speech with features that can be controlled using a simple text prompt (e.g. gender, background noise, speaking rate, pitch and reverberation).

By default, Parler-TTS generates 🎲 random voice characteristics. To ensure 🎯 speaker consistency across generations, try to use consistent descriptions in your prompts.

Note: you do not need to specify the nationality of the speaker in the description (do: "a male speaker", don't: "a french male speaker")

""" ) with gr.Row(): with gr.Column(): input_text = gr.Textbox(label="Input Text", lines=2, value=default_text, elem_id="input_text") description = gr.Textbox(label="Description", lines=2, value=default_description, elem_id="input_description") run_button = gr.Button("Generate Audio", variant="primary") with gr.Column(): audio_out = gr.Audio(label="Parler-TTS generation", type="numpy", elem_id="audio_out") inputs = [input_text, description] outputs = [audio_out] run_button.click(fn=gen_tts, inputs=inputs, outputs=outputs, queue=True) gr.Examples(examples=examples, fn=gen_tts, inputs=inputs, outputs=outputs, cache_examples=True) gr.HTML( """

Tips for ensuring good generation:

""" ) block.queue() block.launch(share=True)