app.py

import logging
import sys
import gradio as gr
from transformers import pipeline, AutoModelForCTC, Wav2Vec2Processor, Wav2Vec2ProcessorWithLM

logging.basicConfig(
    format="%(asctime)s - %(levelname)s - %(name)s -   %(message)s",
    datefmt="%m/%d/%Y %H:%M:%S",
    handlers=[logging.StreamHandler(sys.stdout)],
)
logger = logging.getLogger(__name__)
logger.setLevel(logging.DEBUG)


LARGE_MODEL_BY_LANGUAGE = {
    "Arabic": {"model_id": "jonatasgrosman/wav2vec2-large-xlsr-53-arabic", "has_lm": False},
    "Chinese": {"model_id": "jonatasgrosman/wav2vec2-large-xlsr-53-chinese-zh-cn", "has_lm": False},
    "Dutch": {"model_id": "jonatasgrosman/wav2vec2-large-xlsr-53-dutch", "has_lm": True},
    "English": {"model_id": "jonatasgrosman/wav2vec2-large-xlsr-53-english", "has_lm": True},
    "Finnish": {"model_id": "jonatasgrosman/wav2vec2-large-xlsr-53-finnish", "has_lm": False},
    "French": {"model_id": "jonatasgrosman/wav2vec2-large-xlsr-53-french", "has_lm": True},
    "German": {"model_id": "jonatasgrosman/wav2vec2-large-xlsr-53-german", "has_lm": True},
    "Greek": {"model_id": "jonatasgrosman/wav2vec2-large-xlsr-53-greek", "has_lm": False},
    "Hungarian": {"model_id": "jonatasgrosman/wav2vec2-large-xlsr-53-hungarian", "has_lm": False},
    "Italian": {"model_id": "jonatasgrosman/wav2vec2-large-xlsr-53-italian", "has_lm": True},
    "Japanese": {"model_id": "jonatasgrosman/wav2vec2-large-xlsr-53-japanese", "has_lm": False},
    "Persian": {"model_id": "jonatasgrosman/wav2vec2-large-xlsr-53-persian", "has_lm": False},
    "Polish": {"model_id": "jonatasgrosman/wav2vec2-large-xlsr-53-polish", "has_lm": True},
    "Portuguese": {"model_id": "jonatasgrosman/wav2vec2-large-xlsr-53-portuguese", "has_lm": True},
    "Russian": {"model_id": "jonatasgrosman/wav2vec2-large-xlsr-53-russian", "has_lm": True},
    "Spanish": {"model_id": "jonatasgrosman/wav2vec2-large-xlsr-53-spanish", "has_lm": True},
}

XLARGE_MODEL_BY_LANGUAGE = {
    "Dutch": {"model_id": "jonatasgrosman/wav2vec2-xls-r-1b-dutch", "has_lm": True},
    "English": {"model_id": "jonatasgrosman/wav2vec2-xls-r-1b-english", "has_lm": True},
    "French": {"model_id": "jonatasgrosman/wav2vec2-xls-r-1b-french", "has_lm": True},
    "German": {"model_id": "jonatasgrosman/wav2vec2-xls-r-1b-german", "has_lm": True},
    "Italian": {"model_id": "jonatasgrosman/wav2vec2-xls-r-1b-italian", "has_lm": True},
    "Polish": {"model_id": "jonatasgrosman/wav2vec2-xls-r-1b-polish", "has_lm": True},
    "Portuguese": {"model_id": "jonatasgrosman/wav2vec2-xls-r-1b-portuguese", "has_lm": True},
    "Russian": {"model_id": "jonatasgrosman/wav2vec2-xls-r-1b-russian", "has_lm": True},
    "Spanish": {"model_id": "jonatasgrosman/wav2vec2-xls-r-1b-spanish", "has_lm": True},
}


# LANGUAGES = sorted(LARGE_MODEL_BY_LANGUAGE.keys())

# the container given by HF has 16GB of RAM, so we need to limit the number of models to load
LANGUAGES = sorted(XLARGE_MODEL_BY_LANGUAGE.keys())
CACHED_MODELS_BY_ID = {}


def run(input_file, language, decoding_type, history, model_size="300M"):

    logger.info(f"Running ASR {language}-{model_size}-{decoding_type} for {input_file}")

    history = history or []

    if model_size == "300M":
        model = LARGE_MODEL_BY_LANGUAGE.get(language, None)
    else:
        model = XLARGE_MODEL_BY_LANGUAGE.get(language, None)

    if model is None:
        history.append({
            "error_message": f"Model size {model_size} not found for {language} language :("
        })
    elif decoding_type == "LM" and not model["has_lm"]:
        history.append({
            "error_message": f"LM not available for {language} language :("
        })
    else:

        # model_instance = AutoModelForCTC.from_pretrained(model["model_id"])
        model_instance = CACHED_MODELS_BY_ID.get(model["model_id"], None)
        if model_instance is None:
            model_instance = AutoModelForCTC.from_pretrained(model["model_id"])
            CACHED_MODELS_BY_ID[model["model_id"]] = model_instance

        if decoding_type == "LM":
            processor = Wav2Vec2ProcessorWithLM.from_pretrained(model["model_id"])
            asr = pipeline("automatic-speech-recognition", model=model_instance, tokenizer=processor.tokenizer, 
                           feature_extractor=processor.feature_extractor, decoder=processor.decoder)
        else:
            processor = Wav2Vec2Processor.from_pretrained(model["model_id"])
            asr = pipeline("automatic-speech-recognition", model=model_instance, tokenizer=processor.tokenizer, 
                           feature_extractor=processor.feature_extractor, decoder=None)

        transcription = asr(input_file, chunk_length_s=5, stride_length_s=1)["text"]

        logger.info(f"Transcription for {input_file}: {transcription}")

        history.append({
            "model_id": model["model_id"],
            "language": language,
            "model_size": model_size,
            "decoding_type": decoding_type,
            "transcription": transcription,
            "error_message": None
        })

    html_output = "<div class='result'>"
    for item in history:
        if item["error_message"] is not None:
            html_output += f"<div class='result_item result_item_error'>{item['error_message']}</div>"
        else:
            url_suffix = " + LM" if item["decoding_type"] == "LM" else ""
            html_output += "<div class='result_item result_item_success'>"
            html_output += f'<strong><a target="_blank" href="https://huggingface.co/{item["model_id"]}">{item["model_id"]}{url_suffix}</a></strong><br/><br/>'
            html_output += f'{item["transcription"]}<br/>'
            html_output += "</div>"
    html_output += "</div>"

    return html_output, history


gr.Interface(
    run,
    inputs=[
        gr.inputs.Audio(source="microphone", type="filepath", label="Record something..."),
        gr.inputs.Radio(label="Language", choices=LANGUAGES),
        gr.inputs.Radio(label="Decoding type", choices=["greedy", "LM"]),
        # gr.inputs.Radio(label="Model size", choices=["300M", "1B"]),
        "state"
    ],
    outputs=[
        gr.outputs.HTML(label="Outputs"),
        "state"
    ],
    title="Automatic Speech Recognition",
    description="",
    css="""
    .result {display:flex;flex-direction:column}
    .result_item {padding:15px;margin-bottom:8px;border-radius:15px;width:100%}
    .result_item_success {background-color:mediumaquamarine;color:white;align-self:start}
    .result_item_error {background-color:#ff7070;color:white;align-self:start}
    """,
    allow_screenshot=False,
    allow_flagging="never",
    theme="grass"
).launch(enable_queue=True)