Mlx Whisper

Local speech-to-text with MLX Whisper (Apple Silicon optimized, no API key).

Security Scan

The skill's requirements and runtime instructions are consistent with a local Apple-Silicon speech-to-text wrapper — nothing requested or instructed appears disproportionate to that purpose.

Purpose & Capability

Name/description say local MLX Whisper for Apple Silicon; SKILL.md requires the mlx_whisper binary and documents usage and models. The declared install hint (pip package 'mlx-whisper') matches the stated capability. No unrelated credentials, binaries, or config paths are requested.

Instruction Scope

Instructions are simple command examples invoking the local mlx_whisper binary on audio/video files, and note where models cache (~/.cache/huggingface/). They do not instruct reading unrelated system files, exfiltrating data, or accessing unrelated env vars.

Install Mechanism

The SKILL.md contains an install metadata entry recommending 'pip install mlx-whisper'. This is a normal distribution mechanism for Python CLIs, but it means code will be installed from PyPI (moderate trust requirement) and the binary will run locally. Model files are downloaded on first use (network activity) — expected but worth noting.

Credentials

No environment variables, credentials, or config paths are required by the skill beyond the documented model cache path. That is proportionate to a local transcription tool that downloads public models.

Persistence & Privilege

Skill is not always-enabled and is user-invocable; it does not request persistent system-level changes or modify other skills' configurations. Autonomous invocation is allowed (platform default) but there are no additional privilege escalations requested.

Assessment

This skill is coherent for local speech-to-text on Apple Silicon, but before installing: (1) confirm you trust the 'mlx-whisper' PyPI package and its maintainer, (2) expect large model downloads (~100MB–3GB) to ~/.cache/huggingface/ (ensure disk space), and (3) run installs in a virtualenv or isolated environment if you want to limit risk. Network activity to download models is expected; if you require offline-only operation, verify models are pre-downloaded and trusted. If you need stronger assurances, review the upstream project source (the GitHub link) before installing/running the binary.

Local speech-to-text using Apple MLX, optimized for Apple Silicon Macs.

Quick Start

mlx_whisper /path/to/audio.mp3 --model mlx-community/whisper-large-v3-turbo

Common Usage

# Transcribe to text file
mlx_whisper audio.m4a -f txt -o ./output

# Transcribe with language hint
mlx_whisper audio.mp3 --language en --model mlx-community/whisper-large-v3-turbo

# Generate subtitles (SRT)
mlx_whisper video.mp4 -f srt -o ./subs

# Translate to English
mlx_whisper foreign.mp3 --task translate

Models (download on first use)

Model	Size	Speed	Quality
mlx-community/whisper-tiny	~75MB	Fastest	Basic
mlx-community/whisper-base	~140MB	Fast	Good
mlx-community/whisper-small	~470MB	Medium	Better
mlx-community/whisper-medium	~1.5GB	Slower	Great
mlx-community/whisper-large-v3	~3GB	Slowest	Best
mlx-community/whisper-large-v3-turbo	~1.6GB	Fast	Excellent (Recommended)

Notes

• Requires Apple Silicon Mac (M1/M2/M3/M4)
• Models cache to ~/.cache/huggingface/
• Default model is mlx-community/whisper-tiny; use --model mlx-community/whisper-large-v3-turbo for best results