Skip to main content

voxtype — local voice dictation that types where your cursor is

voxtype turns speech into text in whatever app has focus — your editor, browser, terminal, or a coding agent's prompt box. Everything runs on your machine: no cloud, no accounts, no audio leaving your Mac.

macOS only (Intel and Apple Silicon). voxtype relies on macOS APIs for typing (accessibility), hotkey suppression (the CGEvent tap), and its overlay (AppKit). It is not supported on Linux or Windows — the toggle hotkey can't be suppressed there (it leaks into your text), there is no overlay, and it is untested. On Apple Silicon (M1 or later) it defaults to the parakeet-mlx Apple-GPU engine; Intel Macs use the parakeet CPU engine.

Install

uv tool install voxtype

That's it. The install includes the best engine for your machine — the Apple-GPU parakeet-mlx engine on Apple Silicon (best accuracy AND speed), the CPU parakeet engine everywhere else — and that engine is the runtime default too, so no flags are needed.

Or run without installing: uvx voxtype. (No uv? pipx install voxtype works too.)

Why it's different

  • One key to dictate. Hit the toggle hotkey (default Ctrl+;), speak as long as you like, hit it again — the whole take is transcribed with full sentence context and typed in one go (a single edit to undo, in most apps). No chopping at pauses, sub-second results.
  • Or go hands-free. Say "hey claude" (configurable, with alias spellings) to start dictating; say "stop listening" or go quiet to re-arm. The hotkey still works as an override.
  • Spoken submit. Say "go", "over", or "submit" as a standalone utterance and voxtype presses Enter — dictate a prompt to a coding agent and send it without touching the keyboard.
  • Clean transcripts. Standalone fillers (uh, um, ...) are stripped and 3+ word stutters collapse ("I I I think" becomes "I think") with on-device regex. Optionally (post_process = true), a small local LLM, "S1-mini" by "Superwhisper" run by llama.cpp, also adds punctuation and drops self-corrections. Or use the opt-in Cerebras provider (gpt-oss-120b over its API; transcripts leave your Mac), which also splits long dictations into paragraphs and lists.
  • Visible when it matters. A floating waveform pill appears ONLY while recording — red waves as you speak. Pill on screen = mic live; no pill = not listening.
  • Nothing is ever lost. Dictated into the wrong window? A paste-again hotkey re-types the last session's transcript at the cursor; optional clipboard mirroring too. Escape cancels a recording outright.

Quick start

  1. Install (see above), then grant your terminal Microphone, Accessibility, and Input Monitoring permissions (System Settings → Privacy & Security) when prompted.

  2. Start dictating:

    voxtype --segmentation hold
    

    Press Ctrl+;, speak (watch the waveform pill), press Ctrl+; again — the whole take types at your cursor. Esc cancels a take.

  3. Make it yours: run voxtype setup for an interactive walkthrough — it asks a handful of explained questions (engine, mode, hotkey, ...) and writes your ~/.config/voxtype/config.toml. (Prefer editing by hand? voxtype init drops a fully commented sample instead.) After that, plain voxtype starts with your settings.

Hotkey & voice vocabulary

Action Default Configurable as
Toggle recording Ctrl+; hotkey
Cancel recording (discard) Esc (only while recording) cancel_hotkey
Re-type last transcript off paste_hotkey
Submit (press Enter) say "go" / "over" / "submit" submit_phrases
Stop dictating say "stop listening" stop_phrase

Not sure how to spell a chord? Run voxtype hotkey, press the combo you want, and it prints the exact line to put in your config (e.g. hotkey = "<ctrl>+;").

Engines

Engine Availability Notes
parakeet-mlx included (default on Apple Silicon) Parakeet-TDT 0.6b v3 on the Apple GPU via MLX: fp16 accuracy at ~40x realtime — best accuracy AND speed
parakeet included (default elsewhere) Parakeet-TDT 0.6b v3 on CPU via sherpa-onnx (~490 MB download)
moonshine opt-in: voxtype[moonshine] Streaming Moonshine models, smallest footprint (no Intel-macOS build)

Agent skill (Claude Code + Codex)

voxtype skill installs a guide skill into both Claude Code and Codex (from the cctools-plugins marketplace), so you can just ask your agent "help me install voxtype" and it walks you through everything:

voxtype skill

It sets up whichever of the two agents is on your machine and skips the other.

Documentation

Full reference — every config key with its default, the engine comparison, and all CLI flags: https://pchalasani.github.io/claude-code-tools/tools/voxtype/

History

voxtype previously shipped as voice-type inside claude-code-tools. It still reads a legacy ~/.config/voice-type/config.toml and reuses the old ~/.cache/voice-type model cache, so upgrading is seamless.

MIT license.

Metadata

Release files for voxtype 0.3.3

For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.

Source distribution (sdist)

Source distribution for voxtype 0.3.3
File Size Uploaded Provenance
voxtype-0.3.3.tar.gz 151.1 kB Without provenance Details

Built distribution (wheel)

Table of built distributions (wheels) for voxtype 0.3.3
File Interpreter ABI Platform
voxtype-0.3.3-py3-none-any.whl Python 3 none any Details

Total release size: 258.4 kB

Release files / voxtype-0.3.3.tar.gz

Download URL voxtype-0.3.3.tar.gz
Size 151.1 kB
Tags Source
SHA-256 checksum
How to use checksums
15b43621c998b08242d570b31decddc27447cbc9e68888fcf54e856bccf3b4bd
BLAKE2b-256 checksum
How to use checksums
0b734d566f4fab7ac45ffb234aeb58720f46a26a25b99d70989f1069e30eef94
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
No
Uploaded via uv/0.5.8

Release files / voxtype-0.3.3-py3-none-any.whl

Download URL voxtype-0.3.3-py3-none-any.whl
Size 107.3 kB
Tags Python 3
SHA-256 checksum
How to use checksums
8c9f8d5a61edebfc793284249f0cb455ef8b14be4a206d610babfa051743afe5
BLAKE2b-256 checksum
How to use checksums
5628a994f22cb26019b9ee68c6c647284e0a66e873bb0f500532a82ba399dd7c
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
No
Uploaded via uv/0.5.8

Security for voxtype 0.3.3

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

No provenance found

None of the files in this release were uploaded with provenance metadata.

A note on package security

PyPI is a repository for distributing packages, not a service for auditing them. While third-party security researchers and our staff do their best to detect and remove malicious packages, PyPI cannot guarantee the quality, security or suitability of any given package.

Anthropic, PBC Visionary sponsor Bloomberg Visionary sponsor Hudson River Trading Visionary sponsor Meta Visionary sponsor NVIDIA Visionary sponsor Microsoft Sustainability sponsor Depot Continuous Integration AWS Cloud computing and Security Sponsor Datadog Monitoring Fastly CDN Google Download Analytics Sentry Error logging StatusPage Status page