# Parakeet MLX

An implementation of Nvidia's Parakeet automatic speech recognition models for Apple Silicon using MLX, with a command-line transcription tool.

Parakeet MLX ports Nvidia's Parakeet automatic speech recognition models to run on Apple Silicon through the MLX framework. Its command-line tool transcribes one or more audio files into text, SRT, VTT or JSON output, with configurable output directories and filename templates and an option to include word-level timestamps in SRT or VTT output.

Decoding can run in greedy or beam mode, with beam decoding currently limited to TDT models and configurable through parameters such as beam size, length penalty, patience and a duration reward; long audio is split into overlapping chunks with configurable chunk and overlap durations. The model used can be swapped by pointing the tool at a different Hugging Face repository.

It installs as a Python package through uv or pip, needs ffmpeg on the system for the command-line tool to work, and is released under the Apache-2.0 license.

Website: https://github.com/senstella/parakeet-mlx