Skip to content

voice-setup.md

Latest commit

 

History

History
76 lines (55 loc) · 2.39 KB

voice-setup.md

File metadata and controls

76 lines (55 loc) · 2.39 KB

Voice Setup

Voice input requires sox (audio capture) and whisper.cpp (local transcription).

  • arm64 Homebrew (/opt/homebrew/) required on Apple Silicon — x86 Homebrew runs through Rosetta without Metal GPU (10-20x slower)
  • Binaries needed: rec (from sox), whisper-cli and whisper-server (from whisper-cpp)
  • Whisper model: ~/.local/share/whisper-cpp/ or Homebrew share dir — large-v3-turbo recommended (~1.5GB)
  • GPU memory: ~1.8GB (shared across sessions, one whisper-server instance)

Apple Silicon (M1/M2/M3/M4)

Important: You must use arm64 Homebrew (/opt/homebrew/). The x86 Homebrew (/usr/local/) installs Intel binaries that run through Rosetta 2 without Metal GPU — transcription will be 10-20x slower.

# Check your Homebrew architecture
brew --prefix
# /opt/homebrew  → arm64 (correct)
# /usr/local     → x86 (need to install arm64 Homebrew)

If you only have x86 Homebrew:

# Install arm64 Homebrew (coexists with x86, doesn't affect it)
arch -arm64 /bin/bash -c "$(curl -fsSL https://raw.githubusercontent.com/Homebrew/install/HEAD/install.sh)"

# Add to your shell profile (~/.zshrc)
eval "$(/opt/homebrew/bin/brew shellenv)"

Install with arm64 Homebrew:

/opt/homebrew/bin/brew install sox whisper-cpp

Intel Mac

brew install sox whisper-cpp

Download Whisper Model

whisper-cli --download-model large-v3-turbo   # ~1.5GB, best quality/speed balance

Models are saved to ~/.local/share/whisper-cpp/. The bridge auto-selects the best available model:

Model Size Speed (M1 Max, Metal) Accuracy Best for
large-v3-turbo 1.5GB ~3-5s for 10s audio Excellent Recommended for Apple Silicon
small 466MB ~2-3s Good Limited disk space
base 148MB ~1-2s Fair Fallback (auto-selected if no Metal)

Verify Setup

# Check binary is arm64 with Metal (Apple Silicon)
file $(which whisper-cli)
# → Mach-O 64-bit executable arm64  ← correct

otool -L $(which whisper-cli) | grep metal
# → libggml-metal.0.dylib  ← Metal GPU enabled

The bridge auto-detects Metal support at startup and logs:

[Voice] whisper-cli: arm64=true, metal=true (/opt/homebrew/bin/whisper-cli)
[Voice] Selected whisper model: ~/.local/share/whisper-cpp/ggml-large-v3-turbo.bin