@acpfx/stt-kyutai

Local speech-to-text via Kyutai moshi. Runs entirely on-device -- no API key needed. Uses MLX on macOS (Apple Silicon) and PyTorch on Linux/Windows (CUDA when available).

Usage

This package is a pipeline node for @acpfx/cli. See the CLI package for installation and usage.

Requires uv for Python dependency management. GPU acceleration is auto-detected at runtime.

Manifest

• Consumes: audio.chunk
• Emits: speech.partial, speech.final, speech.pause, lifecycle.ready, lifecycle.done, log

Settings

| Name | Type | Default | Description | |------|------|---------|-------------| | model | string | kyutai/stt-1b-en_fr-candle | HuggingFace model ID | | language | string | en | Language code | | device | string | auto | Compute device: auto, cpu, cuda, or metal |

Pipeline Example

nodes:
  mic:
    use: "@acpfx/mic-speaker"
    settings: { speaker: player }
    outputs: [stt]
  stt:
    use: "@acpfx/stt-kyutai"
    settings: { language: en, device: auto }
    outputs: [bridge]

GPU Acceleration

| Platform | Acceleration | How | |----------|-------------|-----| | macOS | MLX (Apple Silicon) | Automatic, with int8 quantization | | Linux/Windows | CUDA (PyTorch) | Automatic when NVIDIA GPU detected | | All | CPU | Fallback, works everywhere |

External Links

• Kyutai -- Open-weight AI research lab
• Delayed Streams Modeling -- Kyutai's streaming speech model architecture

License

ISC