Sea Shell

Local speech-to-text that runs entirely on your Mac. No cloud, no API keys, no latency.

Uses whisper.cpp with Metal GPU acceleration for fast transcription.

Features

• Always listening - Auto-detects when you start/stop speaking
• Never misses speech - Concurrent architecture transcribes while still listening
• Transcribe audio files - Drag-and-drop a file onto the window, press F for a file picker, or run seashell file.mp3 — wav, mp3, ogg, flac and more (auto-converted)
• 30-second chunking - Long recordings are automatically split for faster transcription
• Pause/resume - Space to pause, transcribes captured audio before pausing
• Copy to clipboard - Press C to copy transcript
• GPU accelerated - Uses Apple Metal for fast inference

Requirements

• macOS (Apple Silicon recommended)
• Bun - JavaScript runtime
• sox - Audio recording (brew install sox)
• cmake - For building whisper.cpp (brew install cmake)
• git - For cloning whisper.cpp

Installation

# Clone the repo
git clone https://github.com/stupart/seashell.git
cd seashell

# Run the installer
chmod +x install.sh
./install.sh

The installer will:

1. Build whisper.cpp with Metal support
2. Download the Whisper large-v3-turbo model (547MB)
3. Download the Silero VAD model
4. Install dependencies
5. Create global seashell command

Usage

seashell

Controls

Key	Action
Space	Pause/Resume
F	Transcribe an audio file (opens a file picker)
C	Copy transcript to clipboard
Delete	Clear transcript
Q or Esc	Quit

Transcribe a file

Got an existing recording? seashell transcribes files too — no mic required:

seashell interview.m4a              # prints the transcript to stdout
seashell voicmemo.mp3 meeting.wav   # transcribe several, in order

Supported out of the box: wav, mp3, ogg, flac — anything else is auto-converted via afconvert first. Because it writes to stdout, it pipes: seashell talk.mp3 > talk.txt.

Or do it live from inside the TUI: drag an audio file from Finder onto the window (its path pastes in and transcribes), or press F for a native file picker. The mic pauses while the file transcribes, shows progress, and resumes listening when it's done.

How It Works

1. sox listens for voice activity (1.5% threshold)
2. When speech is detected, recording begins
3. After 2 seconds of silence (or 30 seconds max), recording stops
4. whisper.cpp transcribes the audio using GPU
5. A new listener starts immediately (concurrent with transcription)
6. Transcribed text appears in the terminal

Models

• Whisper large-v3-turbo-q5_0 - Main transcription model (547MB, quantized)
• Silero VAD v6.2.0 - Voice activity detection

License

MIT

---

Sea Shell is the local-first STT engine behind conch, a hands-free voice loop for Claude Code.

A small open experiment from Blueprint Studio — we build AI products that feel good to use.