app.py

import torch

from transformers import pipeline

import numpy as np
import gradio as gr

def _grab_best_device(use_gpu=True):
    if torch.cuda.device_count() > 0 and use_gpu:
        device = "cuda"
    else:
        device = "cpu"
    return device

device = _grab_best_device()

default_model_per_language = {
    "marathi": "facebook/mms-tts-mar"
}

models_per_language = {
    "marathi": ["ylacombe/mms-mar-finetuned-monospeaker"]
}

HUB_PATH = "ylacombe/vits_ljs_midlands_male_monospeaker"


pipe_dict = {
    "current_model": "ylacombe/vits_ljs_midlands_male_monospeaker",
    "pipe":  pipeline("text-to-speech", model=HUB_PATH, device=0),
    "original_pipe": pipeline("text-to-speech", model=default_model_per_language["marathi"], device=0),
    "language": "english",
}

title =      """
# Explore MMS finetuning
## Or how to access truely multilingual TTS
Massively Multilingual Speech (MMS) models are light-weight, low-latency TTS models based on the [VITS architecture](https://huggingface.co/docs/transformers/model_doc/vits).
Meta's [MMS](https://arxiv.org/abs/2305.13516) project, aiming to provide speech technology across a diverse range of languages. You can find more details about the supported languages and their ISO 639-3 codes in the [MMS Language Coverage Overview](https://dl.fbaipublicfiles.com/mms/misc/language_coverage_mms.html),
and see all MMS-TTS checkpoints on the Hugging Face Hub: [facebook/mms-tts](https://huggingface.co/models?sort=trending&search=facebook%2Fmms-tts).
    
Coupled with the right data and the right training recipe, you can get an excellent finetuned version of every MMS checkpoints in **20 minutes** with as little as **80 to 150 samples**.            
Training recipe available in this [github repository](https://github.com/ylacombe/finetune-hf-vits)!
            """

max_speakers = 1


# Inference
def generate_audio(text, model_id, language):

    if pipe_dict["language"] != language:
        gr.Warning(f"Language has changed - loading new default model: {default_model_per_language[language]}")
        pipe_dict["language"] = language
        pipe_dict["original_pipe"] = pipeline("text-to-speech", model=default_model_per_language[language], device=0)
    
    if pipe_dict["current_model"] != model_id:
        gr.Warning("Model has changed - loading new model")
        pipe_dict["pipe"] = pipeline("text-to-speech", model=model_id, device=0)
        pipe_dict["current_model"] = model_id

    num_speakers = pipe_dict["pipe"].model.config.num_speakers

    out = []
    # first generate original model result
    output = pipe_dict["original_pipe"](text)
    output =  gr.Audio(value = (output["sampling_rate"], output["audio"].squeeze()), type="numpy", autoplay=False, label=f"Non finetuned model prediction {default_model_per_language[language]}", show_label=True,
                               visible=True)
    out.append(output)
    
    
    if num_speakers>1:
        for i in range(min(num_speakers, max_speakers - 1)):
            forward_params = {"speaker_id": i}
            output = pipe_dict["pipe"](text, forward_params=forward_params)
            
            output =  gr.Audio(value = (output["sampling_rate"], output["audio"].squeeze()), type="numpy", autoplay=False, label=f"Generated Audio - speaker {i}", show_label=True,
                               visible=True)
            out.append(output)
        out.extend([gr.Audio(visible=False)]*(max_speakers-num_speakers))
    else:
        output = pipe_dict["pipe"](text)
        output =  gr.Audio(value = (output["sampling_rate"], output["audio"].squeeze()), type="numpy", autoplay=False, label="Generated Audio - Mono speaker", show_label=True,
                               visible=True)
        out.append(output)
        out.extend([gr.Audio(visible=False)]*(max_speakers-2))
    return out


css = """
#container{
    margin: 0 auto;
    max-width: 80rem;
}
#intro{
    max-width: 100%;
    text-align: center;
    margin: 0 auto;
}
"""
# Gradio blocks demo    
with gr.Blocks(css=css) as demo_blocks:
    gr.Markdown(title, elem_id="intro")

    with gr.Row():
        with gr.Column():
            inp_text = gr.Textbox(label="Input Text", info="What sentence would you like to synthesise?")
            btn = gr.Button("Generate Audio!")
            language = gr.Dropdown(
                default_model_per_language.keys(),
                value = "marathi",
                label = "language",
                info = "Language that you want to test"
            )
            
            model_id = gr.Dropdown(
                    models_per_language["marathi"],
                    value="ylacombe/mms-mar-finetuned-monospeaker", 
                    label="Model", 
                    info="Model you want to test",
                    )
                
        with gr.Column():
            outputs = []
            for i in range(max_speakers):
                out_audio = gr.Audio(type="numpy", autoplay=False, label=f"Generated Audio - speaker {i}", show_label=True, visible=False)
                outputs.append(out_audio)

    with gr.Accordion("Datasets and models details", open=False):
        gr.Markdown("""
        
For each language, we used 100 to 150 samples of a single speaker to finetune the model.
### Spanish
* **Model**: [Spanish MMS TTS](https://huggingface.co/facebook/mms-tts-spa).
* **Datasets**:
    - [Chilean Spanish TTS dataset](https://huggingface.co/datasets/ylacombe/google-chilean-spanish).
### Tamil
* **Model**: [Tamil MMS TTS](https://huggingface.co/facebook/mms-tts-tam).
* **Datasets**:
    - [Tamil TTS dataset](https://huggingface.co/datasets/ylacombe/google-tamil).
### Gujarati
* **Model**: [Gujarati MMS TTS](https://huggingface.co/facebook/mms-tts-guj).
* **Datasets**:
    - [Gujarati TTS dataset](https://huggingface.co/datasets/ylacombe/google-gujarati).
### Marathi
* **Model**: [Marathi MMS TTS](https://huggingface.co/facebook/mms-tts-mar).
* **Datasets**:
    - [Marathi TTS dataset](https://huggingface.co/datasets/ylacombe/google-chilean-marathi).
### English
* **Model**: [VITS-ljs](https://huggingface.co/kakao-enterprise/vits-ljs)
* **Dataset**: [British Isles Accent](https://huggingface.co/datasets/ylacombe/english_dialects). For each accent, we used 100 to 150 samples of a single speaker to finetune [VITS-ljs](https://huggingface.co/kakao-enterprise/vits-ljs).
                    
                    """) 

    with gr.Accordion("Run VITS and MMS with transformers", open=False):
        gr.Markdown(
            """
        ```bash
        pip install transformers
        ```
        ```py
        from transformers import pipeline
        import scipy
        pipe = pipeline("text-to-speech", model="kakao-enterprise/vits-ljs", device=0)
        
        results = pipe("A cinematic shot of a baby racoon wearing an intricate italian priest robe")
        # write to a wav file
        scipy.io.wavfile.write("audio_vits.wav", rate=results["sampling_rate"], data=results["audio"].squeeze())
        ```
        """
        )


    language.change(lambda language: gr.Dropdown(
                    models_per_language[language],
                    value=models_per_language[language][0], 
                    label="Model", 
                    info="Model you want to test",
                    ),
                    language,
                    model_id
                   )
    
    btn.click(generate_audio, [inp_text, model_id, language], outputs)
    

demo_blocks.queue().launch()