app.py

from datetime import datetime, timedelta
import os

import gradio as gr
import nemo.collections.asr as nemo_asr
import wandb

MODEL_HISTORY_DAYS = 180
WANDB_ENTITY = os.environ.get("WANDB_ENTITY", "tarteel")
WANDB_PROJECT_NAME = os.environ.get("WANDB_PROJECT_NAME", "nemo-experiments")

wandb_api = wandb.Api(overrides={"entity": WANDB_ENTITY})

all_artifacts_versions = [
    version
    for version in [
        collection.versions()
        for collection in wandb_api.artifact_type(
            type_name="model", project=WANDB_PROJECT_NAME
        ).collections()
    ]
]

latest_artifacts = [
    artifact
    for artifact_versions in all_artifacts_versions
    for artifact in artifact_versions
    if (
        datetime.fromisoformat(artifact.created_at)
        > datetime.now() - timedelta(days=MODEL_HISTORY_DAYS)  # last 180 days
        and artifact.state != "DELETED"
    )
]
latest_artifacts.sort(key=lambda a: a.created_at, reverse=True)

models = {artifact.name: None for artifact in latest_artifacts}


def lazy_load_models(models_names):
    for model_name in models_names:
        model = models[model_name]
        if not model:
            models[model_name] = nemo_asr.models.ASRModel.restore_from(
                list(filter(lambda x: x.name == model_name, latest_artifacts))[0].file()
            )
            models[model_name].eval()


def transcribe(audio_mic, audio_file, models_names):
    lazy_load_models(models_names)
    # transcribe audio_mic and audio_file separately
    # because transcribe() fails is path is empty
    transcription_mic = "\n".join(
        [
            f"{model_name} => {models[model_name].transcribe([audio_mic])[0]}"
            for model_name in models_names
        ]
        if audio_mic
        else ""
    )
    transcription_file = "\n".join(
        [
            f"{model_name} => {models[model_name].transcribe([audio_file])[0]}"
            for model_name in models_names
        ]
        if audio_file
        else ""
    )
    return transcription_mic, transcription_file


model_selection = list(models.keys())

demo = gr.Blocks()

with demo:
    gr.Markdown(
        """
        # ﷽
        These are the latest* Tarteel models.

        Please note that the models are lazy loaded.
        This means that the first time you use a model,
        it might take some time to be downloaded and loaded for inference.

        *: last 180 days since the space was launched.
        To update the list, restart the space.
        """
    )
    with gr.Row():
        audio_mic = gr.Audio(source="microphone", type="filepath", label="Microphone")
        audio_file = gr.Audio(source="upload", type="filepath", label="File")

    with gr.Row():
        output_mic = gr.TextArea(label="Microphone Transcription")
        output_file = gr.TextArea(label="Audio Transcription")

    models_names = gr.CheckboxGroup(model_selection, label="Select Models to Use")

    b1 = gr.Button("Transcribe")

    b1.click(
        transcribe,
        inputs=[audio_mic, audio_file, models_names],
        outputs=[output_mic, output_file],
    )

demo.launch()