mirror of
https://github.com/stablyai/orca.git
synced 2026-11-19 04:23:57 +00:00
* fix(speech): drop a canceled dictation's queued decodes, also during Finish A cancel during "Finishing…" never reached the speech process, and a cancel that did waited in line behind every decode queued before it, so the next dictation sat behind audio nobody wanted. The host now sends the discard during Finish too, and the speech process answers it at once and skips the stream's queued work. * fix(speech): close an audio port whose stream was canceled before it attached * fix(speech): restart the speech process when a canceled decode outlasts a short grace A cancel now answers at once, so a decode already running for the canceled stream (hung in native code, or just long) could hold the queue for the next dictation. The host waits up to 2 s for the speech process to reach the discard; past that, with no other dictation on it, the process restarts, and a dictation starting meanwhile waits for that instead of reusing it. * fix(speech): skip a canceled stream's remaining chunks inside one feed or finish * fix(speech): keep a failed restart after a cancel from failing the next dictation