Sherpa-onnx Kokoro streaming-method audit

Kokoro-Hexagon 0a03be39Updated 2026-09-26

Source was already mirrored in a local upstream mirror; no clone or checkout mutation was needed. This audit reads upstream commit 040afe360a38e25daaa325ce8889abf93ea02609 by immutable Git object. It is a scheduling comparison only. No sherpa-onnx code, ONNX model, runtime, or frontend is a Kokoro-Hexagon product input.

What the callback actually means

sherpa-onnx/csrc/offline-tts-kokoro-impl.h:243-263 converts text to token-ID units. piper-phonemize-lexicon.cc:570-595 iterates the phonemizer's sentence groups; :235-269 adds boundary IDs and further splits a long group by the model's maximum token count.
offline-tts-kokoro-impl.h:267-286 forces Kokoro batch size one even when a larger sentence setting is supplied. At :308-325, each unit is passed to Process, appended to the full answer, and only then handed to the callback. At :420-466, Process makes one complete ONNX Runtime model call and copies its full output. Thus the callback is per completed sentence/token-limited unit, not frame streaming within one Kokoro run.
dotnet-examples/kokoro-tts-play/Program.cs:77-93,163-175 starts an audio stream, copies each delivered unit to a playback queue, and plays while later units synthesize. Its queue is not capacity-bounded, and its playback callback allocates/copies arrays at :117-159. These are example choices, not a latency or robustness standard for our Android audio path.

Kokoro-Hexagon consequence

Use completed legal utterance units as one optional scheduling mode after the exact full-utterance phoneme-to-PCM path works. Keep the model and AAudio stream resident; hand PCM to a bounded playback queue with cancellation and explicit drain. Never call a completed-unit callback intra-model streaming. Compare short-first-unit and steady larger-unit schedules against exact one-shot execution on identical admitted text, voice, speed, and weights. Kokoro's context and full-span AdaIN statistics make a segment cut potentially audible, so segmentation is not stock-equivalent to the one-shot utterance. Preserve the one-shot path and gate any segmented mode by measured first-audio latency, gaps/underruns, memory, and listening results.

This was a source audit, not a benchmark. No sherpa-onnx timing, Kokoro-Hexagon speech timing, device playback, or quality result is inferred from it.