Inventor(s)

Abstract

While a push-to-talk key is held, a background worker re-recognises the most recent audio about once a second and shows the interim text in a live draft card above the recording HUD. Drafts and the final pass share one on-device speech model and one inference lock. A key release during a draft pass would therefore make the final transcription wait for the whole pass, around a second on long utterances. Here the final pass increments a waiting counter before it queues for the lock. A draft is skipped outright if a final is waiting or the lock is busy, so drafts never queue. A draft that is already running decodes token by token with the same model and prompt as the final, and checks the counter before every token. When a final is waiting it stops and returns empty text, never a partial result. Tokens are decoded to text only once at the end, because one token can be half of a multi-byte character. The final pass keeps its unchanged one-shot decoding call. On an 11.9 second utterance whose draft pass took 1.25 seconds, the running draft gave way within 0.46 seconds.

Creative Commons License

Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.

Share

COinS