v0.0.14

Highlights

End-of-turn detection (Smart Turn v3.2)

• Added OnnxSmartTurn, an ONNX Runtime wrapper for Pipecat Smart Turn v3.2, and a TurnCompletionInterface so any end-of-turn classifier can plug into turn detection. The published graph takes the last 8 s of 16 kHz audio and returns the probability that the user has finished; the Whisper log-mel front-end is embedded, so there is no host-side feature extraction.
• TurnDetector now holds a turn when the classifier vetoes a VAD pause: audio keeps accumulating, resumed speech continues the same turn, and turn_completion_max_silence ends it anyway. Eager STT respects the veto. Configure with AgentConfig::turn_completion_threshold and turn_completion_max_silence.
• C API: sc_turn_completion_vtable_t, sc_pipeline_set_turn_completion(), and the two matching sc_config_t fields, so Swift and Kotlin hosts can supply their own classifier the same way as the VAD.
• New speech turn <utterance.wav> CLI (speech_smart_turn), model downloads for soniqo/Smart-Turn-v3.2-ONNX, and docs including the README mirrors. Closes #145.

Desktop packages

• Linux amd64/arm64: .deb and .tar.gz CLI/server packages
• Windows x64: self-contained .zip with ONNX Runtime and PowerShell model downloader
• OpenAI-compatible local audio: POST /v1/audio/speech and POST /v1/audio/transcriptions

SpeechCore XCFramework

Checksum: 68517f044b98a54e8779dccab850ae383ad5a5595471430c9e10118abec345cd

Usage in Package.swift

.binaryTarget(
    name: "CSpeechCore",
    url: "https://github.com/soniqo/speech-core/releases/download/v0.0.14/SpeechCore.xcframework.zip",
    checksum: "68517f044b98a54e8779dccab850ae383ad5a5595471430c9e10118abec345cd"
)

What's Changed

• Add Smart Turn v3.2 end-of-turn classifier to turn detection by @ivan-digital in #146

Full Changelog: v0.0.13...v0.0.14