app.py

from pathlib import Path

import gradio as gr

from transformers import pipeline

DEFAULT_MODEL = "ginic/data_seed_bs64_4_wav2vec2-large-xlsr-53-buckeye-ipa"


VALID_MODELS = [
    "ctaguchi/wav2vec2-large-xlsr-japlmthufielta-ipa-plus-2000",
    "ginic/data_seed_bs64_1_wav2vec2-large-xlsr-53-buckeye-ipa",
    "ginic/data_seed_bs64_2_wav2vec2-large-xlsr-53-buckeye-ipa",
    "ginic/data_seed_bs64_3_wav2vec2-large-xlsr-53-buckeye-ipa",
    "ginic/data_seed_bs64_4_wav2vec2-large-xlsr-53-buckeye-ipa",
    "ginic/gender_split_30_female_1_wav2vec2-large-xlsr-53-buckeye-ipa",
    "ginic/gender_split_30_female_2_wav2vec2-large-xlsr-53-buckeye-ipa",
    "ginic/gender_split_30_female_3_wav2vec2-large-xlsr-53-buckeye-ipa",
    "ginic/gender_split_30_female_4_wav2vec2-large-xlsr-53-buckeye-ipa",
    "ginic/gender_split_30_female_5_wav2vec2-large-xlsr-53-buckeye-ipa",
    "ginic/gender_split_70_female_1_wav2vec2-large-xlsr-53-buckeye-ipa",
    "ginic/gender_split_70_female_2_wav2vec2-large-xlsr-53-buckeye-ipa",
    "ginic/gender_split_70_female_3_wav2vec2-large-xlsr-53-buckeye-ipa",
    "ginic/gender_split_70_female_4_wav2vec2-large-xlsr-53-buckeye-ipa",
    "ginic/gender_split_70_female_5_wav2vec2-large-xlsr-53-buckeye-ipa",
]


def load_model_and_predict(model_name: str, audio_in: str, model_state: dict):
    if model_state["model_name"] != model_name:
        model_state = {
            "loaded_model": pipeline(
                task="automatic-speech-recognition", model=model_name
            ),
            "model_name": model_name,
        }

    return (
        model_state["loaded_model"](audio_in)["text"],
        model_state,
        gr.DownloadButton("Download TextGrid file", visible=True),
    )


def download_textgrid(audio_in, textgrid_tier_name, prediction):
    # TODO
    pass


def launch_demo():
    initial_model = {
        "loaded_model": pipeline(
            task="automatic-speech-recognition", model=DEFAULT_MODEL
        ),
        "model_name": DEFAULT_MODEL,
    }

    with gr.Blocks() as demo:
        gr.Markdown(
            """# Automatic International Phonetic Alphabet Transcription
            This demo allows you to experiment with producing phonetic transcriptions of uploaded or recorded audio using a selected automatic speech recognition (ASR) model.""",
        )
        model_name = gr.Dropdown(
            VALID_MODELS,
            value=DEFAULT_MODEL,
            label="IPA transcription ASR model",
            info="Select the model to use for prediction.",
        )
        audio_in = gr.Audio(type="filepath", show_download_button=True)
        model_state = gr.State(value=initial_model)

        prediction = gr.Textbox(label="Predicted IPA transcription")

        textgrid_tier = gr.Textbox(
            label="TextGrid Tier Name", value="transcription", interactive=True
        )

        download_btn = gr.DownloadButton("Download TextGrid file", visible=False)

        # If user updates model name or audio, run prediction
        audio_in.input(
            fn=load_model_and_predict,
            inputs=[model_name, audio_in, model_state],
            outputs=[prediction, model_state, download_btn],
        )
        model_name.change(
            fn=load_model_and_predict,
            inputs=[model_name, audio_in, model_state],
            outputs=[prediction, model_state, download_btn],
        )

    # demo = gr.Interface(
    #     fn=load_model_and_predict,
    #     inputs=[
    #         gr.Dropdown(
    #             VALID_MODELS,
    #             value=DEFAULT_MODEL,
    #             label="IPA transcription ASR model",
    #             info="Select the model to use for prediction.",
    #         ),
    #         gr.Audio(type="filepath", show_download_button=True),
    #         gr.State(
    #             value=initial_model
    #         ),  # Store the name of the currently loaded model
    #     ],
    #     outputs=[gr.Textbox(label="Predicted IPA transcription"), gr.State()],
    #     allow_flagging="never",
    #     title="Automatic International Phonetic Alphabet Transcription",
    #     description="This demo allows you to experiment with producing phonetic transcriptions of uploaded or recorded audio using a selected automatic speech recognition (ASR) model.",
    # )

    demo.launch()


if __name__ == "__main__":
    launch_demo()