senstella

parakeet-mlx: Transcribe Audio on Apple Silicon

parakeet-mlx runs Nvidia Parakeet speech recognition models on Apple Silicon using MLX. It provides command-line and Python interfaces for transcription, timestamps, chunking, and streaming audio.

View on GitHub Compare
Stars
986
Forks
61
Stars, last 30 days
+10

Our take

Health score

45/100

How it is scored
Activity
32
Community
29
Issues
74
Pull requests
46

+10 stars in the last 30 days

  • Closes issues faster than 82% of the projects we track

Project history

Daily GitHub snapshots since Oct 7, 2026. History covers stars and forks; health signals are not tracked historically.

Selected snapshot

Stars

997

Forks

63

Change since Nov 16, 2026: 0 stars·+1 forks

Overview

parakeet-mlx implements Nvidia's Parakeet automatic speech recognition models for Apple Silicon with the MLX framework. It turns audio files or incoming audio chunks into text, with sentence and token timestamps available for alignment.

It suits Mac users who want to run transcription locally through a CLI or integrate it into Python applications. The project supports several Parakeet model variants and configurable decoding, output formats, and attention behavior.

Key Features

  • Transcribe audio files using Parakeet models hosted on Hugging Face.
  • Use either a command-line interface or a Python API.
  • Export transcripts as TXT, SRT, VTT, JSON, or all supported formats.
  • Return sentence- and token-level timestamps, including word highlighting in subtitle output.
  • Process long recordings in overlapping chunks or use local attention to reduce intermediate memory use.
  • Stream audio and inspect current transcription results as chunks arrive.
  • Configure greedy or beam decoding, with beam decoding available for TDT models.

Use Cases

  • Mac users can transcribe recordings without building a custom speech-recognition pipeline.
  • Researchers and teams can create timestamped subtitles from audio files for review or editing.
  • Developers can add speech transcription to Python workflows using the model API and structured alignment results.
  • Applications that receive audio incrementally can use the streaming API to update transcription as chunks arrive.

Getting Started

Install with uv, then transcribe an audio file:

uv add parakeet-mlx -U
parakeet-mlx audio.mp3

The CLI requires ffmpeg to be installed. See the README for Python API examples and configuration options.

Considerations

  • This implementation targets Apple Silicon and uses MLX, so it is not a general-purpose cross-platform transcription package.
  • The CLI depends on ffmpeg for audio handling.
  • Model weights are downloaded from Hugging Face and cached locally; first use requires access to the selected model.
  • Beam decoding is currently limited to TDT models. Chunking and local-attention settings may need tuning for long recordings and available memory.

Found this useful?

Share it with someone who would like parakeet-mlx.