pipecrab-stt: the speech-to-text interface.

[StreamingTranscriber] is the STT capability the conversation pipeline drives: f32 audio a window at a time in, [SttEvent]s out — partial hypotheses while the user is still speaking, then a final transcript, which is what a low-latency conversation loop needs. Concrete models stay behind it, so the pipeline never names one.

An engine reaches that interface one of two ways:

• A native streaming engine (e.g. a streaming Zipformer) implements [StreamingTranscriber] directly, emitting real partials.
• A chunk-final engine (e.g. Moonshine) implements the simpler one-shot [Transcriber] — f32 samples in, one transcript out, no partials — and the [Buffered] adapter lifts it to [StreamingTranscriber] by accumulating the utterance and transcribing it once at the end. So a partial-less engine still plugs into the same streaming interface, without the pipeline knowing the difference.

[SttStage] adapts a [StreamingTranscriber] into a pipeline Stage as a stateless protocol adapter.

Format authority

Format authority flows from the party that knows the requirement, not from whatever the wire happens to carry. Stages reject fatally mismatched formats.

Platform-neutral and wasm32-checkable: the concrete engines live elsewhere (native ort, browser Transformers.js in a Web Worker), each behind these traits, so the interface itself carries no backend dependency and compiles for both the host and wasm32-unknown-unknown.