speech

Engine traits (Vad, SttEngine, TtsEngine), CancelFlag, EngineInfo, and the pieces that sit between the audio path and an engine.

Module	What it does	Verified
endpoint	Endpointer (pure logic over frame probabilities) and UtteranceSegmenter (adds pre-roll and the audio). 300 ms pre-roll, 200 ms minimum speech, 400 to 900 ms end silence (default 600), forced end at 30 s.	Table-driven unit tests
stt_worker	SttWorker: engine on its own thread, bounded input queue (default 4, a full queue hands the job back and is counted), warm-up after load, one terminal event per utterance, per-job and shutdown cancellation.	Unit tests with a fake engine
tts_worker	TtsWorker: engine on its own thread over sentence strings, bounded input queue (default 8) and bounded audio output (default 8, the worker waits rather than drops), TtsTurn numbers sentences and cancels a whole reply. The sentence chunker lives in crates/tutor-engine/src/chunker.rs and feeds this worker.	Unit tests with a fake engine
sherpa_config	Which files each sherpa-onnx model family needs, and the checks that run before the native library loads them.	Unit tests
sherpa (feature sherpa)	SherpaVad, SherpaStt, SherpaTts over sherpa-onnx 1.13.8.	UNVERIFIED

The sherpa feature is UNVERIFIED

It is off by default. The build script of sherpa-onnx-sys downloads the native library from GitHub releases, which the build container cannot reach. The adapters were compile-checked with SHERPA_ONNX_LIB_DIR pointing at an empty folder, which makes cargo check and cargo clippy pass without linking. They have never been linked, loaded a model or produced a transcript or a sound.

Known limits, read from the 1.13.8 source:

• The wrapper's VAD reports detected(), not a probability per frame. SherpaVad returns 0.0 or 1.0, so the endpointer's hysteresis has nothing to work on and sherpa's own minimum durations add to ours. A real probability needs Silero through ort. The owner decides after the F2 run.
• The wrapper panics on a NUL byte in a path or in text. Configurations are validated first and NULs are removed from text.
• An offline decode is one native call that cannot be interrupted. Cancellation is checked before and after it.
• The static libraries sherpa-onnx links include espeak-ng (GPL-3.0). See docs/LICENSE_REGISTER.md section 6.

On Windows with the model files, the owner runs cargo test -p speech --features sherpa (there are no tests for the adapters yet: they need real models) and a small program that loads each adapter, feeds a WAV file and prints the transcript and the first-audio time.