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



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


# input
inputs = [
    gr.inputs.Audio(source="microphone", type="filepath", label='Input audio'),
    gr.Slider(0, 1, value=0, label="Verbatimicity - from non-verbatim (0) to verbatim (1)", info="Choose a transcription style between non-verbatim and verbatim"),
]


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

text = "ASR Transcription Opensource Demo"

# 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
      "
)



interface = gr.Interface(
    fn=recognition,
    inputs=inputs,
    outputs=output,
    title=text,
    description=description,
    theme='huggingface',
)

interface.launch(enable_queue=True)