|
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": "ylacombe/mms-mar-finetuned-monospeaker" |
|
} |
|
|
|
models_per_language = { |
|
"marathi": ["ylacombe/mms-mar-finetuned-monospeaker"] |
|
} |
|
|
|
HUB_PATH = "ylacombe/mms-mar-finetuned-monospeaker" |
|
|
|
|
|
pipe_dict = { |
|
"current_model": "ylacombe/mms-mar-finetuned-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": "marathi", |
|
} |
|
|
|
title = """ |
|
Marathi Parkinson Enabler: Speaking is a big challenge during Parakinsons. Patients show slurred speech and cannot communicate effectively. |
|
This is marathi text to speech model for parkinson users who want to communicate in Marathi. |
|
""" |
|
|
|
max_speakers = 1 |
|
|
|
|
|
|
|
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) |
|
|
|
num_speakers = pipe_dict["pipe"].model.config.num_speakers |
|
|
|
out = [] |
|
|
|
output = pipe_dict["original_pipe"](text) |
|
output = gr.Audio(value = (output["sampling_rate"], output["audio"].squeeze()), type="numpy", autoplay=True, label=f"Finetuned model prediction {default_model_per_language[language]}", show_label=True, |
|
visible=True) |
|
|
|
return output |
|
|
|
|
|
css = """ |
|
#container{ |
|
margin: 0 auto; |
|
max-width: 80rem; |
|
} |
|
#intro{ |
|
max-width: 100%; |
|
text-align: center; |
|
margin: 0 auto; |
|
} |
|
""" |
|
|
|
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(): |
|
output = gr.Audio(type="numpy", autoplay=False, label=f"Generated Audio", show_label=True, visible=False) |
|
|
|
with gr.Accordion("Datasets and models details", open=False): |
|
gr.Markdown(""" |
|
|
|
### Marathi |
|
* **Model**: [Marathi MMS TTS](https://huggingface.co/facebook/mms-tts-mar). |
|
* **Datasets**: |
|
- [Marathi TTS dataset](https://huggingface.co/datasets/ylacombe/google-chilean-marathi). |
|
""") |
|
|
|
|
|
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], output) |
|
|
|
|
|
demo_blocks.queue().launch() |
