app.py

from funasr import AutoModel
from funasr.utils.postprocess_utils import rich_transcription_postprocess
from modelscope import snapshot_download

import datetime
import math
import io
import os
import tempfile
import json
from typing import Optional

import torch
import gradio as gr

from config import model_config

device = "cuda:0" if torch.cuda.is_available() else "cpu"
model_dir = snapshot_download(model_config['model_dir'])

model = AutoModel(
    model=model_dir,
    trust_remote_code=False,
    remote_code="./model.py",
    vad_model="fsmn-vad",
    punc_model="ct-punc", 
    vad_kwargs={"max_single_segment_time": 15000},
    ncpu=torch.get_num_threads(),
    batch_size=1,
    hub="hf",
    device=device,
)

def transcribe_audio(file_path, vad_model="fsmn-vad", vad_kwargs='{"max_single_segment_time": 15000}', 
                     batch_size=1, language="auto", use_itn=True, batch_size_s=60, 
                     merge_vad=True, merge_length_s=15, batch_size_threshold_s=50, 
                     hotword=" ", ban_emo_unk=True):
    try:
        vad_kwargs = json.loads(vad_kwargs)
        
        temp_file_path = file_path

        res = model.generate(
            input=temp_file_path,
            cache={},
            language=language,
            use_itn=use_itn,
            batch_size_s=batch_size_s,
            merge_vad=merge_vad,
            merge_length_s=merge_length_s,
            batch_size_threshold_s=batch_size_threshold_s,
            hotword=hotword,
            ban_emo_unk=ban_emo_unk
        )

        return res[0]

    except Exception as e:
        return str(e)

inputs = [
    gr.Audio(type="filepath"),
    gr.Textbox(value="fsmn-vad", label="VAD Model"),
    gr.Textbox(value='{"max_single_segment_time": 15000}', label="VAD Kwargs"),
    gr.Slider(1, 10, value=1, step=1, label="Batch Size"),
    gr.Textbox(value="auto", label="Language"),
    gr.Checkbox(value=True, label="Use ITN"),
    gr.Slider(30, 120, value=60, step=1, label="Batch Size (seconds)"),
    gr.Checkbox(value=True, label="Merge VAD"),
    gr.Slider(5, 60, value=15, step=1, label="Merge Length (seconds)"),
    gr.Slider(10, 100, value=50, step=1, label="Batch Size Threshold (seconds)"),
    gr.Textbox(value=" ", label="Hotword"),
    gr.Checkbox(value=True, label="Ban Emotional Unknown"),
]

outputs = gr.Textbox(label="Transcription")

gr.Interface(
    fn=transcribe_audio, 
    inputs=inputs, 
    outputs=outputs, 
    title="ASR Transcription with FunASR"
).launch()