alphacep/asr

1import logging2import sys3import gradio as gr4import vosk5import json6import subprocess7 8logging.basicConfig(9    format="%(asctime)s - %(levelname)s - %(name)s -   %(message)s",10    datefmt="%m/%d/%Y %H:%M:%S",11    handlers=[logging.StreamHandler(sys.stdout)],12)13logger = logging.getLogger(__name__)14logger.setLevel(logging.DEBUG)15 16LARGE_MODEL_BY_LANGUAGE = {17    "Russian": {"model_id": "vosk-model-ru-0.42"},18    "Chinese": {"model_id": "vosk-model-cn-0.22"},19    "English": {"model_id": "vosk-model-en-us-0.22"},20    "French": {"model_id": "vosk-model-fr-0.22"},21    "German": {"model_id": "vosk-model-de-0.22"},22    "Italian": {"model_id": "vosk-model-it-0.22"},23    "Japanese": {"model_id": "vosk-model-ja-0.22"},24    "Hindi": {"model_id": "vosk-model-hi-0.22"},25    "Persian": {"model_id": "vosk-model-fa-0.5"},26    "Uzbek": {"model_id": "vosk-model-small-uz-0.22"},27}28 29LANGUAGES = sorted(LARGE_MODEL_BY_LANGUAGE.keys())30CACHED_MODELS_BY_ID = {}31 32def asr(model, input_file):33 34    rec = vosk.KaldiRecognizer(model, 16000.0)35    results = []36 37    process = subprocess.Popen(f'ffmpeg -loglevel quiet -i {input_file} -ar 16000 -ac 1 -f s16le -'.split(),38                            stdout=subprocess.PIPE)39 40    while True:41        data = process.stdout.read(4000)42        if len(data) == 0:43            break44        if rec.AcceptWaveform(data):45            jres = json.loads(rec.Result())46            results.append(jres['text'])47 48    jres = json.loads(rec.FinalResult())49    results.append(jres['text'])50 51    return " ".join(results)52 53 54def run(input_file, language, history):55 56    logger.info(f"Running ASR for {language} for {input_file}")57 58    history = history or []59 60    model = LARGE_MODEL_BY_LANGUAGE.get(language, None)61 62    if model is None:63        history.append({64            "error_message": f"Failed to find a model for {language} language :("65        })66    elif input_file is None:67        history.append({68            "error_message": f"Record input audio first"69        })70    else:71        model_instance = CACHED_MODELS_BY_ID.get(model["model_id"], None)72        if model_instance is None:73            model_instance = vosk.Model(model_name=model["model_id"])74            CACHED_MODELS_BY_ID[model["model_id"]] = model_instance75 76        transcription = asr(model_instance, input_file)77 78        logger.info(f"Transcription for {input_file}: {transcription}")79 80        history.append({81            "model_id": model["model_id"],82            "language": language,83            "transcription": transcription,84            "error_message": None85        })86 87    html_output = "<div class='result'>"88    for item in history:89        if item["error_message"] is not None:90            html_output += f"<div class='result_item result_item_error'>{item['error_message']}</div>"91        else:92            html_output += "<div class='result_item result_item_success'>"93            html_output += f'{item["transcription"]}<br/>'94            html_output += "</div>"95    html_output += "</div>"96 97    return html_output, history98 99 100gr.Interface(101    run,102    inputs=[103        gr.inputs.Audio(source="microphone", type="filepath", label="Record something..."),104        gr.inputs.Radio(label="Language", choices=LANGUAGES),105        "state"106    ],107    outputs=[108        gr.outputs.HTML(label="Outputs"),109        "state"110    ],111    title="Automatic Speech Recognition",112    description="",113    css="""114    .result {display:flex;flex-direction:column}115    .result_item {padding:15px;margin-bottom:8px;border-radius:15px;width:100%}116    .result_item_success {background-color:mediumaquamarine;color:white;align-self:start}117    .result_item_error {background-color:#ff7070;color:white;align-self:start}118    """,119    allow_flagging="never",120    theme="default"121).launch(enable_queue=True)122