skadoosh 0.6.4

Modular, low-latency local voice agent framework: VAD → Whisper STT → streaming LLM → ONNX TTS with barge-in
docs.rs failed to build skadoosh-0.6.4
Please check the build logs for more information.
See Builds for ideas on how to fix a failed build, or Metadata for how to configure docs.rs builds.
If you believe this is docs.rs' fault, open an issue.
Visit the last successful build: skadoosh-0.11.0

skadoosh

crates.io docs.rs CI license

A fast, fully local voice agent in Rust. Mic in, Silero VAD listens, Whisper transcribes, your LLM replies, and Kokoro speaks each clause as it lands. Interrupt mid-sentence — it shuts up instantly. No cloud, no fuss. OpenAI-compatible backends also work.

Use it as a binary or as a library with pluggable engines.

Install

cargo add skadoosh          # library
cargo install skadoosh      # binary (needs cmake, clang, ALSA dev headers)

5-line SDK

use skadoosh::{Agent, Config, Result};

fn main() -> Result<()> {
    Agent::builder().config(Config::default()).build()?.run()
}

That's a full voice agent. See examples/ for text mode, mock engines, and custom backends.

Quickstart (binary)

ollama pull qwen2.5:0.5b && ollama serve
./scripts/download_models.sh           # VAD + Whisper (~76 MB)
./scripts/download_models.sh --with-kokoro  # + Kokoro TTS (~320 MB, needs espeak-ng)
skadoosh

Talk. It answers. Talk over it — it stops.

skadoosh --repl                         # text chat mode
skadoosh --say "Hello." --out-wav hi.wav
skadoosh --selftest tests/data/jfk.wav  # headless end-to-end test

Key features

  • Clause-level streaming — TTS starts on the first clause, not the full reply
  • Real barge-in — speak anytime, old audio stops in ~5–10 ms, lock-free
  • No cloud — everything runs locally; OpenAI-compatible APIs also supported
  • Plug any engine — implement SttEngine, LlmBackend, or TtsEngine and drop them in
  • GPU acceleration — CUDA, CoreML, DirectML, ROCm via Cargo features (gpu-cuda, etc.)
  • Tool calling — LLMs run shell commands, results feed back into conversation
  • Emotion-aware TTS — keyword-driven speed modulation (excited/calm/neutral)
  • Hold music — chord progression plays during long tool executions
  • #![forbid(unsafe_code)] — zero unsafe, rustls everywhere

Configuration

Every flag has a SKADOOSH_ env var. --api-key is never logged.

Flag Default What it does
--llm-url http://localhost:11434/v1 OpenAI-compatible base URL
--llm-model qwen2.5:0.5b Model name
--api-key — Bearer token for hosted providers
--whisper-model models/ggml-tiny.en.bin STT model
--tts-model — Kokoro ONNX (absent → sine mock)
--mock-tts off Force sine-wave TTS instead of Kokoro
--silence-ms 300 How much silence ends a segment
--vad-threshold 0.5 Speech probability threshold
--push-to-talk off Keyboard toggle for recording
--wake-word — Only process speech containing this
--tts-voice af Kokoro voice key
--tts-speed 1.0 Playback speed (0.5–2.0)
--whisper-model-size tiny Whisper model size for API backends
--hold-music off Chord progression during tool execution
--tts-emotion off Emotion-aware speech speed
--selftest <wav> — Headless end-to-end test

License

Dual-licensed under MIT or Apache-2.0, at your option.