Offline dictation · macOS · Apple Silicon

Speak messy.
Land clean.

Hold Right-Cmd, speak, release — a quick chirp becomes clean text. Fillers, false starts, even a slip into Mandarin become clean English — pasted right where your cursor sits. Every millisecond of audio stays on your Mac.

Install it How it works

~2s from release to clean text on an M1. No cloud, no account, no telemetry.

What it does

Small footprint, opinionated behavior.

Private

100% local

Whisper large-v3-turbo runs quantized on Metal, on-device. Audio is transcribed on your Mac and never touches a network.

Anywhere

Hold to talk

Hold Right-Cmd, speak, release. Text pastes at your cursor in any app; your previous clipboard is restored a moment later.

Bilingual

English + 中文 → English

Speak either or mix. Qwen2.5-3B via Ollama cleans fillers and translates. Prefer instant? A rules-only mode skips the model.

Context

Knows your words

A personal dictionary keeps project names spelled right. Tone adapts to the app — casual in chat, verbatim in terminals.

Honest numbers on a MacBook Air M1 (8GB): about 2.1s from key-release to pasted text for English, ~4.2s with Mandarin auto-detect. Smart mode adds ~0.6s of LLM cleanup for a typical sentence (first Smart dictation after idle pays a few seconds of model load). The model loads on first use and unloads after 10 minutes idle to give your RAM back.

Get it running

One clone, one script, one command.

$ git clone https://github.com/jcoder121/chirp && cd chirp && ./scripts/setup.sh && make run

setup.sh fetches whisper-cpp and the model (~550 MB) and, for Smart mode, optionally installs Ollama + qwen2.5:3b. On first launch, grant Microphone and Accessibility in System Settings, then relaunch. That's the whole ceremony.