Parakeet MLX – Speech‑to‑Text on Apple Silicon

What it is – A Python package that runs Nvidia’s Parakeet automatic‑speech‑recognition (ASR) models on Apple‑silicon Macs using the MLX inference framework. It provides a command‑line tool and a Python API for turning audio files (or live audio streams) into transcriptions with timestamps.

---

Key features (from the README)

Feature	Details
Model support	Loads Parakeet variants (TDT, RNNT, CTC, TDTCTC) from Hugging‑Face (mlx-community/parakeet‑tdt‑0.6b‑v3 etc.).
Apple‑silicon acceleration	Uses the MLX library, which is optimized for the M‑series GPU/CPU.
CLI	parakeet-mlx <audio_files> [OPTIONS] – batch transcription, selectable output formats (txt, srt, vtt, json, or all), word‑level timestamps, chunking for long files, beam‑search decoding, precision control (bf16 / fp32), and local‑attention mode for memory‑efficient long audio.
Python API	from_pretrained loads a model; model.transcribe(path, …) returns an AlignedResult with full text and per‑sentence/ token timestamps.
Decoding options	Greedy (default) or beam search (beam size, length penalty, patience, duration reward).
Chunking & streaming	Split long audio into overlapping chunks; transcribe_stream enables real‑time transcription with configurable attention context and depth.
Local attention	Switch to a sliding‑window attention (rel_pos_local_attn) to reduce memory for very long inputs.
Output templating	Custom filename patterns using {filename}, {parent}, {index}, {date}.
Installation	Via uv (recommended) or pip; requires system ffmpeg.

---

How to use it

CLI quick start

# basic transcription
parakeet-mlx audio.mp3

# VTT subtitles with word‑level timestamps
parakeet-mlx *.mp3 --output-format vtt --highlight-words

# generate every supported format
parakeet-mlx audio.mp3 --output-format all

Python example

from parakeet_mlx import from_pretrained, DecodingConfig, Beam

model = from_pretrained("mlx-community/parakeet-tdt-0.6b-v3")

# simple greedy decode
res = model.transcribe("audio.wav")
print(res.text)

# beam search with custom parameters
cfg = DecodingConfig(decoding=Beam(beam_size=5, length_penalty=0.013,
                                   patience=3.5, duration_reward=0.67))
res = model.transcribe("audio.wav", decoding_config=cfg)
print(res.sentences)

---

Who might find it useful?

• Developers building voice‑enabled macOS apps who need on‑device transcription without sending data to the cloud.
• Researchers prototyping ASR pipelines on Apple hardware, leveraging Nvidia‑trained Parakeet models.
• Content creators who want fast, local subtitle generation (SRT/VTT) for podcasts or videos.

---

Limitations & TODO (as listed)

• No built‑in word‑accuracy enhancement yet.
• Chunking with continuous context is only partially supported; streaming aims to fill that gap.

---

License

Apache 2.0 – free for commercial and research use.

---

All information above is taken directly from the repository’s README.