app.py

import gradio as gr
import torch
from wenet.cli.model import load_model

import os
from huggingface_hub import login

# Load the API token from the environment variables
api_token = os.getenv('HUGGINGFACE_API_TOKEN')
if not api_token:
    raise ValueError("No Hugging Face API token found. Please set the HUGGING_FACE_API_TOKEN environment variable.")

# Login to Hugging Face Hub
login(token=api_token, add_to_git_credential=True)    

def process_cat_embs(cat_embs):
    device = "cpu"
    cat_embs = torch.tensor(
        [float(c) for c in cat_embs.split(',')]).to(device)
    return cat_embs


def download_rev_models():
    from huggingface_hub import hf_hub_download
    import joblib

    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])
    return model

model = download_rev_models()
    

def recognition(audio, style=0):
    if audio is None:
        return "Input Error! Please enter one audio!"
    

    cat_embs = ','.join([str(s) for s in (style, 1-style)])
    cat_embs = process_cat_embs(cat_embs)
    ans = model.transcribe(audio, cat_embs = cat_embs)

    if ans is None:
        return "ERROR! No text output! Please try again!"
    txt = ans['text']
    txt = txt.replace('▁', ' ')
    return txt


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")
    
text = "ASR Transcription Opensource Demo-CPU"

# description
description = (
    " Opensource Automatic Speech Recognition in English"
    
      "Verbatim Transcript style(1) refers to word to word-to-word transcription of an audio" 
      "Non Verbatim Transcript style(0) refers to just conserving the message of the original audio"
)



iface = gr.Interface(
    fn=recognition,
    inputs=[audio_input, style_slider],
    outputs=output_textbox,
    title=text,
    description=description,
    theme='default',
)

iface.launch()