app.py

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

from transformers import pipeline, AutoTokenizer, VitsModel, VitsTokenizer


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 and speaker embeddings
# processor = SpeechT5Processor.from_pretrained("microsoft/speecht5_tts")

# model = SpeechT5ForTextToSpeech.from_pretrained(
#    "microsoft/speecht5_tts").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)


model_nld = VitsModel.from_pretrained("facebook/mms-tts-ara")
tokenizer = VitsTokenizer.from_pretrained("facebook/mms-tts-ara")


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


def synthesise2(text):
    # inputs = processor(text=text, return_tensors="pt")
    # speech = model.generate_speech(inputs["input_ids"].to(device), speaker_embeddings.to(device), vocoder=vocoder)
    inputs = tokenizer(text, return_tensors="pt")
    with torch.no_grad():
        outputs = model_nld(**inputs)
        #speech = outputs.waveform[0]
        audio_tensor = outputs.audio
        # Convert tensor to numpy array
        audio_array = audio_tensor.numpy()
        # Reshape the audio array to 1D
    audio_array = audio_array.reshape(-1)
    return audio_array


def speech_to_speech_translation(audio):
    translated_text = translate(audio)
    synthesised_speech = synthesise2(translated_text)
    synthesised_speech = np.int16(synthesised_speech * 32767)
    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/microsoft/speecht5_tts) model for text-to-speech:

![Cascaded STST](https://huggingface.co/datasets/huggingface-course/audio-course-images/resolve/main/s2st_cascaded.png "Digram 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()