Voice (alpha — push-to-talk)

The web UI now has a 🎤 button next to the answer textarea. Hold it, speak your answer, release — the transcript drops into the textarea so you can edit before submitting. Default backend is fully local: a whisper.cpp subprocess Rocky shells out to. Nothing leaves the machine.

Setup — one line

bash
curl -fsSL https://raw.githubusercontent.com/NVME-git/rocky/main/scripts/install-whisper.sh | sh

The installer:

  • Detects platform (linux-x64, linux-arm64, macos-arm64, macos-x64)
  • Installs whisper-cli — via brew install whisper-cpp on macOS, building from source on Linux (cmake + a C++ compiler required, ~2-5 min)
  • Pulls ggml-base.en.bin (~142 MB) into ~/.rocky/models/
  • Adds a [voice] block to ~/.config/rocky/config.toml

After it finishes, restart rocky view and the mic button is live.

Two backends

BackendLatencyPrivacySetup
whisper-cpp (default)~1-2 s on CPU✓ fully localinstall-whisper.sh
browser (opt-in)real-time✗ Chrome → Google, Safari → Appleflip a config flag

Browser mode is forbidden when privacy.strict = true — Rocky won’t let you accidentally exfiltrate audio. Enable it explicitly with:

[voice]
provider        = "browser"
browser_consent = true

What ships in v0.2

  • Web UI mic button (push-to-hold, WAV encoded client-side at 16 kHz mono)
  • POST /api/transcribe endpoint that calls the configured STT provider
  • whisper.cpp subprocess invocation with clear errors when the binary or model is missing (the web UI surfaces an inline link to the installer)
  • One-line installer (scripts/install-whisper.sh)

What’s planned for later

  • rocky quiz --voice — fully hands-free CLI session: Rocky speaks the question via OS TTS, captures your answer with cpal + webrtc-vad (700ms silence ends an utterance), evaluates, repeats
  • In-binary whisper-rs build (cargo install --features voice) for users who want one binary, no PATH dependency
  • Real-time streaming transcription (Pattern C in ADR 0006)

There is no wake-word / always-on listening — push-to-talk is the only model. By design.

Disabling voice

[voice]
provider = "off"

Mic button stops working immediately, no rebuild needed.