app.py

# -*- coding: utf-8 -*-
"""app.ipynb

Automatically generated by Colaboratory.

Original file is located at
    https://colab.research.google.com/drive/16MxXQeF3O0htL9eQ61aa6ZxnApGg9TKN
"""

import gradio as gr
import numpy as np
import torch

from transformers import pipeline, VitsModel, VitsTokenizer, FSMTForConditionalGeneration, FSMTTokenizer

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

#eng audio to text transformation
asr_pipe = pipeline("automatic-speech-recognition", model="asapp/sew-d-tiny-100k-ft-ls100h", device=device)

#eng text to rus text translation
mname = "facebook/wmt19-en-ru"
tokenizer = FSMTTokenizer.from_pretrained(mname)
model = FSMTForConditionalGeneration.from_pretrained(mname)

#translation_pipe = pipeline("translation", model="facebook/wmt19-en-ru")

#rus text to rus speech transformation
vits_model = VitsModel.from_pretrained("facebook/mms-tts-rus")
vits_tokenizer = VitsTokenizer.from_pretrained("facebook/mms-tts-rus")

def transform_audio_to_speech_en(audio):
    outputs = asr_pipe(audio, max_new_tokens=256, generate_kwargs={"task": "translate"})
    return outputs["text"]

def translator(text):
  input_ids = tokenizer.encode(text, return_tensors="pt")
  outputs = model.generate(input_ids)
  decoded = tokenizer.decode(outputs[0], skip_special_tokens=True)
  return decoded
  #translated_text = translation_pipe(text)
  #return translated_text[0]['translation_text']

def synthesise(translated_text):
    translated_text = translator(translated_text)
    inputs = vits_tokenizer(translated_text, return_tensors="pt")
    with torch.no_grad():
        speech = vits_model(**inputs).waveform
    return speech.cpu()

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

title = "Cascaded STST"
description = """
Demo for cascaded speech-to-speech translation (STST), mapping from source speech in any language to target speech in English. Demo uses OpenAI's [Whisper Base](https://huggingface.co/openai/whisper-base) model for speech translation, and Microsoft's
[SpeechT5 TTS](https://huggingface.co/microsoft/speecht5_tts) 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(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(type="filepath"),
    outputs=gr.Audio(label="Generated Speech", type="numpy"),
    examples=[["/content/drive/MyDrive/test_2.wav"]],
    title=title,
    description=description,
)

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

demo.launch()