app.py

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

from transformers import (
    SpeechT5ForTextToSpeech, SpeechT5HifiGan, SpeechT5Processor, WhisperForConditionalGeneration, WhisperFeatureExtractor, WhisperTokenizer, pipeline,
    BarkModel, BarkProcessor
)


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


feature_extractor = WhisperFeatureExtractor.from_pretrained("openai/whisper-base")
tokenizer = WhisperTokenizer.from_pretrained("openai/whisper-base", language="french", task="translate")

whisper_model = WhisperForConditionalGeneration.from_pretrained("openai/whisper-base")
forced_decoder_ids = tokenizer.get_decoder_prompt_ids(language="french", task="translate")

#load text-to-speech checkpoint and speaker embeddings
processor = SpeechT5Processor.from_pretrained("Apocalypse-19/speecht5_finetuned_french")

model = SpeechT5ForTextToSpeech.from_pretrained("Apocalypse-19/speecht5_finetuned_french").to(device)
vocoder = SpeechT5HifiGan.from_pretrained("microsoft/speecht5_hifigan").to(device)

embeddings_dataset = load_dataset("Matthijs/cmu-arctic-xvectors", split="validation")
speaker_embeddings = torch.tensor(embeddings_dataset[7306]["xvector"]).unsqueeze(0)


def translate(audio):
    # load speech translation checkpoint
    asr_pipe = pipeline(
        "automatic-speech-recognition", 
        model=whisper_model, 
        feature_extractor=feature_extractor,
        tokenizer=tokenizer,
        device=device
    )
    outputs = asr_pipe(audio, max_new_tokens=256, generate_kwargs={"task": "transcribe", "forced_decoder_ids": forced_decoder_ids})
    return outputs["text"]


def synthesise(text):
    
    # inputs = processor(text, voice_preset="v2/fr_speaker_1")
    # speech = bark_model.generate(**inputs).cpu()
    
    inputs = processor(text=text, return_tensors="pt")
    speech = model.generate_speech(inputs["input_ids"].to(device), speaker_embeddings.to(device), vocoder=vocoder)
    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 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/leo-kwan/speecht5_finetuned_voxpopuli_lt) 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()