app.py

"Задание 3"
import gradio as gr
import numpy as np
import torch
import espeakng
import sentencepiece
from datasets import load_dataset

from transformers import pipeline, MarianMTModel, MarianTokenizer, VitsModel, VitsTokenizer

device = "cuda:0" if torch.cuda.is_available() else "cpu"

import phonemizer
model_wav2vec = 'facebook/wav2vec2-lv-60-espeak-cv-ft'
asr_pipe = pipeline("automatic-speech-recognition", model=model_wav2vec, device=device)

# from speech to text
def translate_audio(audio):
    outputs = asr_pipe(audio, max_new_tokens=256, 
                       generate_kwargs={"task": "translate"})
    return outputs["text"]

# translation
def translate_text(text, from_language, target_language):  #to English -mul en, to Russian - en ru
    model_name = f'Helsinki-NLP/opus-mt-{from_language}-{target_language}'
    tokenizer = MarianTokenizer.from_pretrained(model_name)
    model = MarianMTModel.from_pretrained(model_name)

    inputs = tokenizer.encode(text, return_tensors="pt")
    outputs = model.generate(inputs, num_beams=4, max_length=50, early_stopping=True)
    translated_text = tokenizer.decode(outputs[0], skip_special_tokens=True)
    return translated_text

# load text-to-speech checkpoint
#model = pipeline("text-to-speech", model="voxxer/speecht5_finetuned_commonvoice_ru_translit")
model = VitsModel.from_pretrained("voxxer/speecht5_finetuned_commonvoice_ru_translit")
tokenizer = VitsTokenizer.from_pretrained("facebook/mms-tts-rus")

def synthesise(text):
    translated_text = translate_text(text, 'mul', 'en')
    translated_text = translate_text(translate_text, 'en', 'ru')
    inputs = tokenizer(translated_text, return_tensors="pt")
    input_ids = inputs["input_ids"]
    with torch.no_grad():
        outputs = model(input_ids)
    speech = outputs["waveform"]
    return speech.cpu()


def speech_to_speech_translation(audio):
    text_from_audio = translate_audio(audio)
    print(translated_text)
    synthesised_speech = synthesise(text_from_audio)
    synthesised_speech = (synthesised_speech.numpy() * 32767).astype(np.int16)
    return 16000, synthesised_speech[0]


title = "Cascaded STST"
description = """
* В начале происходит распознавание речи с помощью модели facebook/wav2vec2-lv-60-espeak-cv-ft и на выходе получается текст на любом из 60 языков.
* Затем полученный текст переводится сначала на английский с помощью Helsinki-NLP/opus-mt-mul-en, а потом на русский с помощью Helsinki-NLP/opus-mt-en-ru
* На последнем шаге полученный текст озвучивается с помощью fine-tune-говой версии microsoft/speecht5_tts - voxxer/speecht5_finetuned_commonvoice_ru_translit
Demo for cascaded speech-to-speech translation (STST), mapping from source speech in any language to target speech in Russian. Demo uses facebook/mms-tts-rus model for text-to-speech:
![Cascaded STST](https://huggingface.co/datasets/huggingface-course/audio-course-images/resolve/main/s2st_cascaded.png "Diagram of cascaded speech to speech translation")
"""

demo = gr.Blocks()

mic_translate = gr.Interface(
    fn=speech_to_speech_translation,
    inputs=gr.Audio(source="microphone", type="filepath"),
    outputs=gr.Audio(label="Generated Speech", type="numpy"),
    title=title,
    description=description,
)

file_translate = gr.Interface(
    fn=speech_to_speech_translation,
    inputs=gr.Audio(source="upload", type="filepath"),
    outputs=gr.Audio(label="Generated Speech", type="numpy"),
    examples=[["./example.wav"]],
    title=title,
    description=description,
)

with demo:
    gr.TabbedInterface([mic_translate, file_translate], ["Microphone", "Audio File"])

demo.launch()