app.py

from pathlib import Path
import tempfile

import gradio as gr
import librosa
import tgt.core
import tgt.io3
from transformers import pipeline

TEXTGRID_DIR = tempfile.mkdtemp()
DEFAULT_MODEL = "ginic/data_seed_bs64_4_wav2vec2-large-xlsr-53-buckeye-ipa"
TEXTGRID_DOWNLOAD_TEXT = "Download TextGrid file"
TEXTGRID_NAME_INPUT_LABEL = "TextGrid file name"


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",
    "ginic/vary_individuals_old_only_1_wav2vec2-large-xlsr-53-buckeye-ipa",
    "ginic/vary_individuals_old_only_2_wav2vec2-large-xlsr-53-buckeye-ipa",
    "ginic/vary_individuals_old_only_3_wav2vec2-large-xlsr-53-buckeye-ipa",
    "ginic/vary_individuals_young_only_1_wav2vec2-large-xlsr-53-buckeye-ipa",
    "ginic/vary_individuals_young_only_2_wav2vec2-large-xlsr-53-buckeye-ipa",
    "ginic/vary_individuals_young_only_3_wav2vec2-large-xlsr-53-buckeye-ipa",
]


def load_model_and_predict(
    model_name: str,
    audio_in: str,
    model_state: dict,
):
    if audio_in is None:
        return (
            "",
            model_state,
            gr.Textbox(label=TEXTGRID_NAME_INPUT_LABEL, interactive=False),
        )

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

    prediction = model_state["loaded_model"](audio_in)["text"]
    return (
        prediction,
        model_state,
        gr.Textbox(
            label=TEXTGRID_NAME_INPUT_LABEL,
            interactive=True,
            value=Path(audio_in).with_suffix(".TextGrid").name,
        ),
    )


def get_textgrid_contents(audio_in, textgrid_tier_name, transcription_prediction):
    if audio_in is None or transcription_prediction is None:
        return ""

    duration = librosa.get_duration(path=audio_in)

    annotation = tgt.core.Interval(0, duration, transcription_prediction)
    transcription_tier = tgt.core.IntervalTier(
        start_time=0, end_time=duration, name=textgrid_tier_name
    )
    transcription_tier.add_annotation(annotation)
    textgrid = tgt.core.TextGrid()
    textgrid.add_tier(transcription_tier)
    return tgt.io3.export_to_long_textgrid(textgrid)


def write_textgrid(textgrid_contents, textgrid_filename):
    """Writes the text grid contents to a named file in the temporary directory.
    Returns the path for download.
    """
    textgrid_path = Path(TEXTGRID_DIR) / Path(textgrid_filename).name
    textgrid_path.write_text(textgrid_contents)
    return textgrid_path


def get_interactive_download_button(textgrid_contents, textgrid_filename):
    return gr.DownloadButton(
        label=TEXTGRID_DOWNLOAD_TEXT,
        variant="primary",
        interactive=True,
        value=write_textgrid(textgrid_contents, textgrid_filename),
    )


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")

        gr.Markdown("""## TextGrid File Options
                    Change these inputs if you'd like to customize and download the transcription in [TextGrid format](https://www.fon.hum.uva.nl/praat/manual/TextGrid_file_formats.html) for Praat.
                    """)
        textgrid_tier = gr.Textbox(
            label="TextGrid Tier Name", value="transcription", interactive=True
        )

        textgrid_filename = gr.Textbox(
            label=TEXTGRID_NAME_INPUT_LABEL, interactive=False
        )

        textgrid_contents = gr.Textbox(
            label="TextGrid Contents",
            value=get_textgrid_contents,
            inputs=[audio_in, textgrid_tier, prediction],
        )

        download_btn = gr.DownloadButton(
            label=TEXTGRID_DOWNLOAD_TEXT,
            interactive=False,  # Don't allow download button to be active until an upload happened
            variant="primary",
        )

        # Update prediction if model or audio changes
        gr.on(
            triggers=[audio_in.input, model_name.change],
            fn=load_model_and_predict,
            inputs=[model_name, audio_in, model_state],
            outputs=[prediction, model_state, textgrid_filename],
        )

        # Download button becomes interactive if user updates audio or textgrid params
        gr.on(
            triggers=[textgrid_contents.change, textgrid_filename.change],
            fn=get_interactive_download_button,
            inputs=[textgrid_contents, textgrid_filename],
            outputs=[download_btn],
        )

    demo.launch(max_file_size="100mb")


if __name__ == "__main__":
    launch_demo()