Get it for macOS

On-device voice dictation for BB's chat composer and `bb voice transcribe`, powered by Hex and macOS 26's built-in Apple Speech (SpeechTranscriber + SpeechAnalyzer) or local GGUF speech models (Parakeet, Whisper, Qwen3-ASR, Cohere Transcribe, Nemotron). Starts a warm embedded `hex service --embedded` daemon when BB loads the plugin, adds a Voice Provider picker directly to the composer microphone right-click menu, transcodes browser microphone clips (`audio/webm`, `audio/mp4`, `audio/ogg`) to 16 kHz mono PCM WAV via `ffmpeg`, and stops the daemon automatically when BB exits.


Overview

View source

Hex Voice

On-device voice dictation for BB's chat composer and bb voice transcribe, powered by Hex and macOS 26's built-in Apple Speech (SpeechTranscriber + SpeechAnalyzer) or local GGUF speech models.

Why Hex Voice

BB's composer microphone button dispatches recorded speech to the active voice AI service. With Hex Voice installed and selected, pressing the microphone button transcribes your speech locally on Apple Silicon in ~100ms with zero cloud calls, zero API keys, and zero recurring cost.

How it works

  1. Registers the hex (Hex (Apple Speech)) AI service via bb.experimental_aiServices.register.
  2. Adds a Voice Provider picker to the microphone right-click menu — right-clicking the composer microphone button displays all configured voice providers right below Input device, so you can switch between Hex (Apple Speech), Codex, and any other configured voice provider on the fly.
  3. Manages a warm hex service --embedded daemon tied to BB's plugin lifecycle (bb.onDispose and a stdin pipe lease), so it starts automatically when BB loads the plugin and exits cleanly when BB quits or reloads.
  4. Transcodes browser microphone recordings (audio/webm, audio/mp4, audio/ogg, etc.) to 16 kHz mono 16-bit PCM WAV via ffmpeg before sending them to Hex's /transcriptions endpoint.
  5. Supports Apple Speech and local GGUF models — defaults to apple_speech (macOS 26 SpeechTranscriber), with support for parakeet_unified_en, parakeet_v2, parakeet_v3, whisper_large_v3_turbo, qwen3_asr06_b, cohere_transcribe, and nemotron_3_5_asr.

Commands

Check service health and installed models:

bb hex status

Set Hex as BB's active voice service:

bb settings ai-services set voice hex

Transcribe an audio file from the CLI:

bb voice transcribe /path/to/audio.wav

Requirements

  • macOS 26+ on Apple Silicon for apple_speech (or macOS with a local GGUF model for Parakeet/Whisper).
  • hex (>=3.2.0) installed on PATH (for example /opt/homebrew/bin/hex or ~/.local/bin/hex) or configured via the plugin's hexBinaryPath setting.
  • ffmpeg installed (for example brew install ffmpeg) to transcode browser MediaRecorder clips (audio/webm / audio/mp4) to 16 kHz mono PCM WAV.

More from Braedon Saunders

Audio PreviewPlays .m4a and other audio files in BB's file panel, including audio/mp4 that the built-in preview refuses.39BallastKnow which thread is eating your RAM before the machine starts paging. Reads vm_stat page classes rather than the free percentage macOS reports, which counts inactive pages as free and stays reassuring while the box holds gigabytes of swap — pressure is graded on headroom, paging rate, compression and the kernel's own verdict, and the banner names the signal that decided. Consumers roll up by process tree and by BB thread, matched through worktree paths, so a 2.4GB row names a thread you can steer instead of six identical node processes. A guard can throttle BB concurrency, ask the hungriest threads to release memory, stop orphaned processes, or hand the problem to an agent — every one of those is off until you turn it on. Nothing is stopped unless nothing can still be waiting on it, and the apply step takes only ids the plan minted, re-checked against the live process table. macOS and Linux, no account or external service.38Muse CodeAdds Meta's Muse Code as a bb provider over its own session protocol: streamed timeline, approvals, subagents, model picker, health, and the plan's own 5-hour and weekly usage limits as Muse reports them.61Provider UsageShows how much of each provider's plan is left — every rate-limit window, percent remaining, and when it resets, including Muse Code's 5-hour and weekly limits — plus a 7/30/90-day token chart covering every provider, from local transcripts and BB's own usage events.100

More in Agents & Providers 73

Add, choose, configure, route, or coordinate who runs a thread.

View all