app.py

import gradio as gr
import logging
import numpy as np
import torch

from transformers import VitsModel, VitsTokenizer, pipeline
from transformers import M2M100ForConditionalGeneration
from tokenization_small100 import SMALL100Tokenizer


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

target_language = "fr"

# load speech translation checkpoint
asr_pipe = pipeline("automatic-speech-recognition", model="bofenghuang/whisper-small-cv11-french", device=device)
translation_model = M2M100ForConditionalGeneration.from_pretrained("alirezamsh/small100")
translation_tokenizer = SMALL100Tokenizer.from_pretrained("alirezamsh/small100", tgt_lang=target_language)

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


def translate(audio):
    outputs = asr_pipe(audio, max_new_tokens=256, generate_kwargs={"task": "translate"})
    eng_text = outputs["text"]
    encoded_eng_text = translation_tokenizer(eng_text, return_tensors="pt")
    generated_tokens = translation_model.generate(**encoded_eng_text)
    translated_text = translation_tokenizer.batch_decode(generated_tokens, skip_special_tokens=True)
    logging.info(f"Translated Text: {translated_text}")
    return translated_text


def synthesise(text):
    inputs = tokenizer(text, return_tensors="pt")
    with torch.no_grad():
        outputs = model(inputs["input_ids"])
    speech = outputs["waveform"][0]
    logging.info(speech)
    return speech.cpu()


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


title = "Cascaded STST"
description = """
Demo for cascaded speech-to-speech translation (STST), mapping from source speech in any language to target speech in French. Demo uses OpenAI's [Whisper Base](https://huggingface.co/openai/whisper-base) model for ASR, the 
[SMaLL-100](https://huggingface.co/alirezamsh/small100) model for text to text translation and Facebook's[MMS TTS-FRA](https://huggingface.co/facebook/mms-tts-fra) for text-to-speech for french:

![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"])

logging.getLogger().setLevel(logging.INFO)
demo.launch()