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


def recognition(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


# Gradio UI Components
inputs = [
    gr.Audio(source="microphone", type="filepath", label='Input audio'),
    gr.Slider(0, 1, value=0, label="Transcription Style", info="Adjust between non-verbatim (0) and verbatim (1) transcription")
]

output = gr.Textbox(label="Output Text")

# UI and Interface
iface = gr.Interface(
    fn=recognition,
    inputs=inputs,
    outputs=output,
    title="Reverb ASR Transcription",
    description="Supports verbatim and non-verbatim transcription styles.",
    article="<p style='text-align: center'><a href='https://rev.com' target='_blank'>Learn more about Rev</a></p>",
    theme='huggingface'
)

iface.launch(enable_queue=True)