What's new // AUDIOBASH

v3.3.1 June 2026 LATEST

Custom dictionary now works on local transcription.

The custom dictionary was silently ignored by the local Whisper and Parakeet paths — configured corrections only applied to the cloud providers. Now spoken-to-written entries apply to local model output too, matching the Gemini, OpenAI, and Anthropic behavior. Regression tests cover both local paths.

Local dictionary

Corrections apply to Whisper and Parakeet

Matches cloud

Same behavior as the cloud providers

Regression-tested

New tests on both local paths

Custom dictionary Local Whisper Parakeet

v3.3.0 June 2026

Linux lands. Transcription hardens.

AudioBash now ships AppImage and .deb downloads, so the voice terminal runs on Linux alongside Windows and macOS. Batch cloud transcription requests moved into the main process instead of the renderer. Global shortcuts register under Wayland, not just X11 — the Alt bindings work on modern Linux desktops.

Linux builds

AppImage and .deb for x64 desktops

Hardened transcription

Batch requests run in the main process

Wayland shortcuts

Global hotkeys under Wayland

Linux AppImage + .deb Main-process batch transcription Wayland

v3.0 March 2026

Split panes that actually work.

The terminal layout system was rebuilt from scratch. A binary tree data model replaces the old fixed grid, letting you split any pane horizontally or vertically, zoom into a single pane, or snap to presets. A terminal pool keeps your shells alive across layout changes — custom tab names and scrollback survive preset cycling. The remote control feature is gone; Claude Code's native /voice mode handles that now.

Binary tree panes

Split, zoom, resize any pane freely

Terminal pool

Shells persist across layout changes

Ready handshake

No more empty panes on preset switch

Pane system SVG toolbar Remote removed Focus highlight

v2.4.1 February 14, 2026

31K lines deleted. Signed builds incoming.

Stripped 31,000 lines of tunnel services, React PWA, and voice bridge from the remote control system. Laid the groundwork for macOS code signing: entitlements plists, a notarization hook, and hardened runtime config. Set up GitHub Sponsors.

-31K lines

Tunnel, PWA, voice bridge removed

Code signing

Entitlements + notarization ready

Apple Silicon works. Electron 39.

The app was crashing immediately on M1/M2/M3 Macs because our build process broke the code signatures on node-pty's native binaries. ARM64 macOS kills unsigned executables on sight. The fix re-signs everything after packaging. While we were at it, we jumped from Electron 35 to 39, which fixes a GPU lag issue on macOS Tahoe where Electron apps were making WindowServer eat 50%+ GPU.

M1 crash fix

Re-sign binaries after chmod

Electron 39

Tahoe GPU lag fixed

Crash-proof startup

Global error handlers added

Apple Silicon Electron 39 Code signing #29

v2.3.0 January 12, 2026

Local Whisper: transcription without API keys

No API key? No problem. Download Whisper.cpp models directly from Settings and run transcription entirely offline. Plus, quick model cycling in the voice overlay and a new console error viewer for debugging.

Local models

Whisper tiny/base/small with one-click download

Model cycling

Click model name in overlay to switch providers

Error viewer

Console errors captured in status bar

Offline transcription Whisper.cpp ElevenLabs WebSocket fix Resizable directory picker

v2.1.1 January 12, 2026

Voice is back

Turns out, the Content Security Policy was silently blocking all your transcription API calls. ElevenLabs, OpenAI, Anthropic - all of them. The fix was one line in the CSP header, but finding it required diving into the browser console during a remote debugging session from my phone.

// The culprit

connect-src 'self' https://api.elevenlabs.io wss://api.elevenlabs.io...

CSP Fix ElevenLabs OpenAI Anthropic

v2.1.0 January 10, 2026 FEATURE

Real-time transcription

ElevenLabs Scribe v2 brought WebSocket streaming, and it changes everything. Instead of waiting 2-3 seconds after you stop talking, your words appear in ~150ms. It feels like the terminal is reading your mind.

~150ms

Transcription latency

VAD

Auto-commit on silence

WebSocket PCM 16kHz Voice Activity Detection

v2.0.0 January 2, 2026 MAJOR

Hello, macOS

AudioBash went cross-platform. Native Apple Silicon builds (M1/M2/M3), Intel support for older Macs, and automatic shell detection so it opens zsh instead of trying to spawn PowerShell on your Mac. The stress test suite expanded to 120+ tests to catch platform-specific edge cases.

Apple Silicon

Windows

120+ tests

Read the macOS announcement

v1.1.0 December 11, 2025

Make it yours

Custom instructions let you tell the AI how to format your commands. Custom vocabulary maps your weird pronunciations to the right words. Audio chimes when Claude Code needs input. 14 keyboard shortcuts. The whole package.

// Vocabulary example

"lawn mow" → "npm"

// Shortcuts

Alt+A, Alt+M, Alt+R...

Custom instructions Vocabulary mapping Audio chimes