# Vella

> Free, open-source offline dictation and transcription (speech-to-text) for Apple Silicon Macs. Parakeet, Whisper, Qwen3 ASR and Nemotron run locally with custom Metal kernels: on an M5 Max, Parakeet v3 Ultra runs 1.9× faster than on stock MLX, on 28% less energy, with no measurable accuracy loss. Vella never uploads audio or transcripts.

Vella is a macOS menu-bar app (macOS 26+, Apple Silicon). Dictation pastes the transcript where you were typing; Streaming types words as they are recognised. Coding agents can transcribe audio files through the `vella` command or an OpenAI-compatible local API.

## Docs

- [Vella page as Markdown](https://tobynoskillson.com/vella/index.md): what it is, speed and energy against stock MLX, every measured number, privacy
- [README](https://github.com/TobyNoSkillSon/Vella#readme): install, models, the app, privacy
- [User guide](https://github.com/TobyNoSkillSon/Vella/blob/main/docs/USAGE.md): every menu item, recovery, troubleshooting
- [Benchmark data](https://github.com/TobyNoSkillSon/Vella/blob/main/Resources/benchmarks.json): every model and precision, WER, speed, energy, memory

## Agents

- [transcribe skill](https://tobynoskillson.com/vella/.well-known/agent-skills/transcribe/SKILL.md): transcribe audio files offline with Vella (text, JSON, SRT, VTT)
- [Benchmark kit](https://github.com/TobyNoSkillSon/Vella/tree/main/Benchmarks): agents measure Vella on another Mac and, with the user's consent, propose the result

## Optional

- [Source](https://github.com/TobyNoSkillSon/Vella)
- [Support the developer](https://github.com/sponsors/TobyNoSkillSon)
- [Verdict](https://github.com/TobyNoSkillSon/Verdict): local text classification for Mac and coding agents, same family