app.py

import gradio as gr
import numpy as np
import torch
from datasets import load_dataset
import librosa
from transformers import SpeechT5ForTextToSpeech, SpeechT5HifiGan, SpeechT5Processor, pipeline
from transformers import WhisperProcessor, WhisperForConditionalGeneration


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)
asr_processor = WhisperProcessor.from_pretrained("openai/whisper-base")
asr_model = WhisperForConditionalGeneration.from_pretrained("openai/whisper-base").to(device)
asr_forced_decoder_ids = asr_processor.get_decoder_prompt_ids(language="dutch", task="transcribe")



# load text-to-speech checkpoint and speaker embeddings
if 0:
    processor = SpeechT5Processor.from_pretrained("sanchit-gandhi/speecht5_tts_vox_nl")

    model = SpeechT5ForTextToSpeech.from_pretrained("sanchit-gandhi/speecht5_tts_vox_nl").to(device)
if 1:
    from transformers import VitsModel, VitsTokenizer
    model = VitsModel.from_pretrained("Matthijs/mms-tts-fra")
    tokenizer = VitsTokenizer.from_pretrained("Matthijs/mms-tts-fra")

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):
    if 0:
        outputs = asr_pipe(audio, max_new_tokens=256, generate_kwargs={"language":"dutch", "task":"transcribe"})
        return outputs["text"]
    else:
        
        x, sr = librosa.load(audio)
        input_features = asr_processor(x, sampling_rate=16000,  return_tensors="pt").input_features
        predicted_ids = asr_model.generate(input_features, forced_decoder_ids=asr_forced_decoder_ids)
        # decode token ids to text
        transcription = asr_processor.batch_decode(predicted_ids, skip_special_tokens=True)
        return transcription



def synthesise(text):
    if 0:
        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()
    if 1:
        inputs = tokenizer(text, return_tensors="pt")
        input_ids = inputs["input_ids"]


        with torch.no_grad():
            outputs = model(input_ids)

        speech = outputs.audio[0]
        return speech.cpu()


def speech_to_speech_translation(audio):
    translated_text = translate(audio)
    print(translated_text)
    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 Dutch. 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(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()