Skip to main content

Voice dictation daemon using NVIDIA Parakeet on Apple Silicon

Project description

🐦 Birdword

Contextual voice dictation for macOS. Powered by NVIDIA Parakeet running locally on Apple Silicon via MLX.

Press a hotkey, speak, and your words are transcribed and pasted into whatever app is focused. A small LLM post-processes the transcription to fix errors, using project-specific context from a BIRDWORD.md file.

Getting started

Requires macOS on Apple Silicon (M1+) and Python 3.10+.

# Run with uvx (no install needed)
uvx birdword

# Or run in the background
uvx birdword start
uvx birdword stop
uvx birdword status

Context-aware correction

The key idea behind birdword is contextual transcription correction. When dictating into Terminal.app, birdword detects the focused tab's working directory and looks for a BIRDWORD.md file up the directory tree. This lets you teach birdword your project's domain:

Context detection works with:

  • Terminal.app — detects the focused tab's shell working directory
  • VS Code / VS Code Insiders — via the Birdword extension, which works with local and remote (SSH) workspaces

Transcription and pasting work in any app.

uvx birdword init

This creates a BIRDWORD.md with the default prompt template. Edit it to add your project's terms, names, and jargon:

---
transcription_model: mlx-community/parakeet-tdt-0.6b-v2
fix_model: mlx-community/Qwen2.5-1.5B-Instruct-4bit
---

Fix transcription errors. Output only the corrected text.

Example 1:
Input: "the java script function isnt working"
Output: "The JavaScript function isn't working."

Example 2:
Input: "check the get ignore file for the repo"
Output: "Check the .gitignore file for the repo."

Example 3:
Input: "we need to refactor the a p i endpoint"
Output: "We need to refactor the API endpoint."

Key terms: MyClass, some_function, PostgreSQL
Names: Alice, Bob

Input: "{{ transcript }}"
Output:

The file is a Jinja template. {{ transcript }} is replaced with the raw transcription. If omitted, the transcript is appended automatically.

The YAML front matter lets you override models per-project. When you dictate into a Terminal tab whose shell is in that directory (or a child), birdword picks up the nearest BIRDWORD.md and uses it.

Hotkeys

Action Default
Toggle recording Right ⌘ + Space
Hold to record Hold Right ⌘ for >1s, release to transcribe

Hotkeys are configurable:

--hold-key KEY       Hold key (default: rcmd). Options: rcmd, lcmd, ralt, lalt, rshift, lshift, rctrl, lctrl
--toggle-key KEY     Toggle key (default: space). Options: space, return, tab, escape

Options

--model MODEL        Transcription model (default: mlx-community/parakeet-tdt-0.6b-v2)
--fix-model MODEL    Post-processor model (default: mlx-community/Qwen2.5-1.5B-Instruct-4bit)
--no-fix             Disable LLM post-processing

Dashboard

Birdword runs a local web dashboard at localhost:7870. Click the bird in the menu bar → Dashboard… to open it.

  • 📝 History — browse all your transcriptions with timestamps, app name, working directory, and duration. See both the original and corrected text.
  • ⚙️ Settings — configure hotkeys, models, and post-processing. Changes take effect immediately without restarting.

You can also view history from the command line:

uvx birdword history

Menu bar

Birdword shows a bird icon in the menu bar:

  • White — idle
  • 🟡 Yellow — connecting mic
  • 🔴 Red — listening
  • Sparkles — transcribing

Permissions

Birdword needs three macOS permissions, granted to your terminal app:

  • 🎤 Microphone — to record your voice
  • 🔐 Accessibility — to paste text and intercept the hotkey
  • ⌨️ Input Monitoring — to detect the global hotkey

Birdword checks these on startup and tells you what's missing.

License

MIT

Project details


Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

birdword-0.3.0.tar.gz (413.4 kB view details)

Uploaded Source

Built Distribution

If you're not sure about the file name format, learn more about wheel file names.

birdword-0.3.0-py3-none-any.whl (32.4 kB view details)

Uploaded Python 3

File details

Details for the file birdword-0.3.0.tar.gz.

File metadata

  • Download URL: birdword-0.3.0.tar.gz
  • Upload date:
  • Size: 413.4 kB
  • Tags: Source
  • Uploaded using Trusted Publishing? Yes
  • Uploaded via: twine/6.1.0 CPython/3.13.7

File hashes

Hashes for birdword-0.3.0.tar.gz
Algorithm Hash digest
SHA256 a1b67591a8a7676530f5e8815151e9180b4da8a8572883c5051080e48cdbdc92
MD5 bb4828e0fd91d6a4ab9215c5566c18d3
BLAKE2b-256 3bb76dadbe2705fe500684872f68d3d8e3bd408ace920a296e9e378d0204ff9d

See more details on using hashes here.

Provenance

The following attestation bundles were made for birdword-0.3.0.tar.gz:

Publisher: main.yaml on tillahoffmann/birdword

Attestations: Values shown here reflect the state when the release was signed and may no longer be current.

File details

Details for the file birdword-0.3.0-py3-none-any.whl.

File metadata

  • Download URL: birdword-0.3.0-py3-none-any.whl
  • Upload date:
  • Size: 32.4 kB
  • Tags: Python 3
  • Uploaded using Trusted Publishing? Yes
  • Uploaded via: twine/6.1.0 CPython/3.13.7

File hashes

Hashes for birdword-0.3.0-py3-none-any.whl
Algorithm Hash digest
SHA256 82743994306c19de03e00f57ba6d1dacebb8531fdec6e88ca20da9e99674b1ea
MD5 49b7163993fc60b8cdef181077c0dc02
BLAKE2b-256 6af1889a9a8d6a18af95073f1ffdfe418580264dafe5719929c0abbe54058f60

See more details on using hashes here.

Provenance

The following attestation bundles were made for birdword-0.3.0-py3-none-any.whl:

Publisher: main.yaml on tillahoffmann/birdword

Attestations: Values shown here reflect the state when the release was signed and may no longer be current.

Supported by

AWS Cloud computing and Security Sponsor Datadog Monitoring Depot Continuous Integration Fastly CDN Google Download Analytics Pingdom Monitoring Sentry Error logging StatusPage Status page