A single recognition request usually has a duration or token cap

What it is

A recognition request is not an infinite audio pipe. Services and on-device models impose a duration or token cap: often tens of seconds to a couple of minutes, or a maximum token count. Past the cap is outside this request’s contract. Dictation and meeting capture designed as infinite streams will hit this hard wall.

Why it happens

Streaming decoder state, attention windows, billing, and abuse protection are accounted per request. Memory and latency grow linearly or worse with length, and the cloud must not let an open microphone idle forever. Endpointing is not a substitute for a cap: a user can keep talking, keep voicing, the detector never seals, and the timer still fires. A post-wake listening window is an interaction timeout; this is the engine’s capacity for one audio span. Both can exist: the window may stop capture first, or capacity may refuse or cut first. The cap is often invisible to people until they are cut off.

Studying it

Use spoken material of increasing duration: around and past the published cap, is the request a success, a cut, or an error? Record whether a warning came first. Compare actual caps on APIs and on-device models with their docs. Short-command corpora never reach the cap.

Where it stops holding

True streaming meeting transcription is sliced into overlapping sliding windows, so there is no visible “one request” wall, but each window still has a length. A never-uploaded on-device engine can have a wider cap and is still memory-bound. If push-to-talk hold time is shorter than the engine cap, finger fatigue arrives first. Calling every long-speech failure a network problem misses a sixty-second ceiling as a product decision.

Applying it

• Warn remaining duration before dictation starts or as the cap nears; do not cut in silence.
• Document seconds or tokens, and use the same cap in the API and the UI.
• Accept with audio past the cap and confirm a comprehensible failure rather than silent data loss.

Related

• Same group: C7.15.2 Past the cap the system may truncate, chunk, or refuse · C7.15.3 Chunking long speech can cut a semantically complete sentence at a boundary · C7.15.4 Length caps directly constrain dictation usefulness
• Adjacent: C7.02 Endpoint Detection · C7.09 Endpoint Detection and End-of-Utterance
• Search: max utterance length · ASR timeout · token limit