app.py

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

from transformers import SpeechT5ForTextToSpeech, SpeechT5HifiGan, pipeline
from transformers import SpeechT5Processor

token = "<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>"
model_identifier = "tugstugi/mongolian-tts-ljspeech"

processor = SpeechT5Processor.from_pretrained(model_identifier, revision="main", token=token)


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

# load speech translation checkpoint
asr_pipe = pipeline("automatic-speech-recognition", model="facebook/wav2vec2-base-960h", device=device)

# load text-to-speech checkpoint and speaker embeddings
processor = SpeechT5Processor.from_pretrained("tugstugi/mongolian-tts-ljspeech")


model = SpeechT5ForTextToSpeech.from_pretrained("ttskit/ttskit-tts-ljspeech").to(device)
vocoder = SpeechT5HifiGan.from_pretrained("ljspeech/vocoder-cryptron").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):
    outputs = asr_pipe(audio, max_new_tokens=256, generate_kwargs={"task": "translate"})
    return outputs["text"]

def synthesise(text):
    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 Facebook's [Wav2Vec 2.0](https://huggingface.co/facebook/wav2vec2-base-960h) model for speech recognition, and a lightweight text-to-speech model ([ttskit/ttskit-tts-ljspeech](https://huggingface.co/ttskit/ttskit-tts-ljspeech)) along with a lightweight vocoder ([ljspeech/vocoder-cryptron](https://huggingface.co/ljspeech/vocoder-cryptron)).
"""

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()