app.py

import gradio as gr
import torch
from wenet.cli.model import load_model
from huggingface_hub import hf_hub_download
import spaces

REPO_ID = "Revai/reverb-asr"
files = ['reverb_asr_v1.jit.zip', 'tk.units.txt']
downloaded_files = [hf_hub_download(repo_id=REPO_ID, filename=f) for f in files]
model = load_model(downloaded_files[0], downloaded_files[1])
    
def process_cat_embs(style):
    device = torch.device("cpu")
    embedding = torch.tensor([style, 1 - style], device=device)
    return embedding


@spaces.GPU
def transcribe_audio(audio, style=0):
    if not audio:
        return "Input Error! Please enter one audio!"
    
    cat_embs = process_cat_embs(f'{style},{1-style}')
    result = model.transcribe(audio, cat_embs=cat_embs)

    if not result or 'text' not in result:
        return "ERROR! No text output! Please try again!"
    
    text_output = result['text'].replace('▁', ' ')
    return text_output



audio_input = gr.Audio(type="filepath", label="Upload or Record Audio")
style_slider = gr.Slider(0, 1, value=0, step=0.1, label="Transcription Style",
                             info="Adjust the transcription style: 0 (casual) to 1 (formal).")
output_textbox = gr.Textbox(label="Transcription Output")
    
description = "This tool transcribes audio using a customizable transcription style ranging from casual to formal. Upload or record an audio file to begin."

iface = gr.Interface(
        fn=transcribe_audio,
        inputs=[audio_input, style_slider],
        outputs=output_textbox,
        title="Audio Transcription with Style Control",
        description=description,
        theme="default"
    )
    

iface.launch()