app.py

import gradio as gr
import numpy as np
import torch
from datasets import load_dataset

from transformers import VitsModel, VitsTokenizer, pipeline


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

# load speech translation checkpoint
asr_pipe = pipeline("automatic-speech-recognition", model="openai/whisper-base", device=device)

# load text-to-speech checkpoint
tokenizer = VitsTokenizer.from_pretrained("Matthijs/mms-tts-deu")

model = VitsModel.from_pretrained("Matthijs/mms-tts-deu")
model.to(device)


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


def synthesize(text):
    input = tokenizer(text, return_tensors="pt")
    with torch.no_grad():
        output = model(input['input_ids'].to(device))
    
    return output.audio[0].cpu()  


target_dtype = np.int16 # output audio file format expected by Gradio
max_range = np.iinfo(target_dtype).max

def speech_to_speech_translation(audio):
    translated_text = translate(audio)
    synthesized_speech = synthesize(translated_text)
    # normalize audio array by dynamic range of target dtype for Gradio
    synthesized_speech = (synthesized_speech.numpy() * max_range).astype(target_dtype)
    return 16000, synthesized_speech


title = "Cascaded STST"
description = """
Demo for cascaded speech-to-speech translation (STST), mapping from source speech in any language to target speech in German. Demo uses OpenAI's [Whisper Base](https://huggingface.co/openai/whisper-base) model for speech translation, and Facebook's
[MMS](https://huggingface.co/facebook/mms-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(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"),
    title=title,
    description=description,
)

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

demo.launch(debug=True)