app.py

import os

os.system("wget https://www.isi.edu/~ulf/uroman/downloads/uroman-v1.2.7.tar.gz")  
os.system("mkdir uroman")
os.system("tar -zxvf ./uroman-v1.2.7.tar.gz -C ./uroman")
os.system("chmod +x ./uroman/bin/uroman.pl")

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

from transformers import SpeechT5ForTextToSpeech, SpeechT5HifiGan, SpeechT5Processor, pipeline


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

# load speech translation checkpoint
asr_pipe = pipeline("automatic-speech-recognition", model="KoRiF/whisper-small-be", device=device)

# load text-to-speech checkpoint and speaker embeddings
processor = SpeechT5Processor.from_pretrained("microsoft/speecht5_tts")

model = SpeechT5ForTextToSpeech.from_pretrained("KoRiF/speecht5_finetuned_common_voice_be").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, transliteration = lambda txt: txt):
    outputs = asr_pipe(audio, max_new_tokens=256, generate_kwargs={"task": "transcribe", "language": "be"})#larusian
    return transliteration(outputs["text"])

import subprocess

def transliterate_text(text, lang_code=None, use_chart=False, use_cache=True):
    command = ['perl', './uroman/bin/uroman.pl']
    if lang_code:
        command.extend(['-l', lang_code])
    if use_chart:
        command.append('--chart')
    if not use_cache:
        command.append('--no-cache')
    
    process = subprocess.Popen(command, stdin=subprocess.PIPE, stdout=subprocess.PIPE, stderr=subprocess.PIPE,
                               universal_newlines=True)
    output, error = process.communicate(input=text)
    if (error):
        print(f"Error: >>> {error}")
    return output.strip()

language = 'bel'
def transliterate(text):
    return transliterate_text(text, language)

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()



target_dtype = np.int16
max_range = np.iinfo(target_dtype).max

def speech_to_speech_translation(audio):
    translated_text = translate(audio, transliterate)#
    synthesised_speech = synthesise(translated_text)
    synthesised_speech = (synthesised_speech.numpy() * max_range).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./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()