Python + UV demo for using mlx_whisper.

demo-mlx-whisper.py

# /// script
# requires-python = ">=3.11"
# dependencies = [
#     "mlx-whisper",
#     "rich",
#     "typer",
# ]
# ///
import json
import typer
import mlx_whisper

from enum import Enum
from pathlib import Path
from rich import print


class ModelChoices(str, Enum):
    tiny = "tiny"
    base = "base"
    small = "small"
    medium = "medium"
    large = "mlx-community/whisper-large-v3-turbo"
    turbo = "mlx-community/whisper-turbo"


def main(
    input_paths: list[str],
    model: ModelChoices = ModelChoices.turbo,
    word_timestamps: bool = False
):
    model_name = model.value.split("/")[-1]
    for input_path in input_paths:
        filename_stem = Path(input_path).stem
        output_path = Path("captions", f"{filename_stem}.{model_name}.json")
        print(f"{output_path=}")
        if not output_path.exists():
            result = mlx_whisper.transcribe(input_path, path_or_hf_repo=model.value, response_format="srt", word_timestamps=word_timestamps)
            # print(result["text"])
            output_path.write_text(json.dumps(result, indent=2))


if __name__ == "__main__":
    typer.run(main)