app.py

"Задание 3"
import gradio as gr
import numpy as np
import torch

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

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

import phonemizer
# variants: 'voidful/wav2vec2-xlsr-multilingual-56'; facebook/wav2vec2-lv-60-espeak-cv-ft, но здесь не загружается библиотека py-espeak-ng
model_wav2vec = 'openai/whisper-small'
asr_pipe = pipeline("automatic-speech-recognition", model=model_wav2vec, device=device)

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

# translation into Russian
def translate_text(text):  
    # to English - mul en, to Russian - en ru
    model_mul_en = pipeline("translation", model = "Helsinki-NLP/opus-mt-mul-en")
    model_en_ru = pipeline("translation", model = "Helsinki-NLP/opus-mt-en-ru")
    translated_text = model_en_ru(model_mul_en(text)[0]['translation_text'])
    return translated_text[0]['translation_text']

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

def synthesise(text):
    translated_text = translate_text(text)
    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)
    synthesised_speech = synthesise(text_from_audio)
    synthesised_speech = (synthesised_speech.numpy() * 32767).astype(np.int16)
    return 16000, synthesised_speech[0]


title = "Cascaded STST. Russian language version"
description = """
* В начале происходит распознавание речи с помощью модели `openai/whisper-small`.
* 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 модели `facebook/mms-tts-rus model`.
  
Demo for cascaded speech-to-speech translation (STST), mapping from source speech in any language to target speech in Russian.  
Demo uses `openai/whisper-small` for speech-to-text and `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()