Chirp

Free, fast, offline voice-to-text for macOS.

Hold fn, speak, release — text appears at your cursor. Works fully offline by default with no accounts, no servers, and no data leaving your machine. Optionally connect cloud providers for higher-accuracy transcription and AI post-processing.

Install

Download

Grab the latest DMG from GitHub Releases, open it, and drag Chirp to Applications.

Homebrew

brew install --cask stefanpenner/chirp/chirp

Build from source

brew install bazelisk
bazel run //:Chirp

Features

• Hold-to-talk — Press fn (configurable) to record, release to transcribe
• Works everywhere — Text is typed at your cursor in any app
• Speculative preview — See partial transcription as you speak
• Fast — Parakeet TDT 0.6b v3 with Silero VAD, all on-device
• AI Modes — Pipeline presets switchable from the menu bar. Ships with "Offline" and "Offline + Fixup" (on-device T5 grammar correction); create your own with cloud STT, LLM post-processing, or any combination
• Cloud providers — Optional OpenAI, Anthropic, and Google APIs (plus any OpenAI-compatible endpoint)
• Speaker verification — Enroll your voice so Chirp only transcribes you, ignoring other speakers
• Noise reduction — Apple Voice Processing for noise suppression, echo cancellation, and automatic gain control
• Menu bar app — Lives in the menu bar, no dock icon
• Auto-updates — Built-in update checks via Sparkle

Setup

Chirp needs two permissions on first launch:

1. Microphone — macOS will prompt automatically
2. Accessibility — Required for typing text at your cursor: System Settings → Privacy & Security → Accessibility, add Chirp and enable the toggle

On first launch, Chirp downloads the speech recognition model (~465 MB). This is a one-time download stored in ~/Library/Application Support/Chirp/.

Updating

Chirp checks for updates automatically. You can also check manually from the menu bar → Check for Updates…

brew upgrade chirp

Requirements

• macOS 15+
• ~700 MB disk space (app + model)

Privacy

In the default offline modes, all processing runs locally. Audio is processed in memory, never recorded or stored. No telemetry, no tracking.

When you use a cloud AI mode, audio or text is sent to the provider you configure. No data is sent unless you explicitly create and activate a cloud mode.

Acknowledgments

• sherpa-onnx — speech recognition engine (Apache 2.0)
• ONNX Runtime — model inference runtime (MIT)
• NVIDIA Parakeet TDT 0.6b v3 — speech-to-text model (CC-BY-4.0)
• Silero VAD — voice activity detection (MIT)
• Sparkle — auto-update framework (MIT)

Full license texts are in THIRD_PARTY_NOTICES.

Development

brew install bazelisk
bazel build //:Chirp           # build
bazel run //:Chirp             # build and launch
bazel test //...               # run tests
bazel run //:package -- 0.3.0  # create signed DMG