After recognition, mark a shift into processing, distinct from still-listening

What it is

After the endpoint, audio often still has to finish decoding, skill routing, and a network round-trip. In that interval the system is no longer listening; it is processing. If the look does not change, people keep talking, the rest of the sentence never enters this turn, or it becomes a new one. Processing must be marked apart from listening.

Why it happens

The state machine leaves capturing for processing at the endpoint. The microphone may already be off, or still open for barge-in, but this turn’s hypothesis is sealed. If the UI still shows a jumping waveform, people think they can still add words. If it goes blank, it looks hung. A processing look—different motion, a “recognizing” label, level animation stopped—announces that capture has ended and the result has not arrived. That can coexist with partials still rewriting: partials are the text layer; processing is the channel layer—new audio is not merged into this turn. It is also not session timeout: timeout closes the whole engagement; processing is a phase inside this turn.

Studying it

Under an artificially long processing delay, compare “still looks like listening,” an explicit processing state, and no feedback. Measure how often people keep talking after the endpoint, whether those words enter the next turn, and whether they knew to stop. Align recordings of endpoint time with UI-switch time.

Where it stops holding

On-device recognition so fast that processing cannot be drawn may go straight to a final. Dialogue systems that allow barge-in may still be listening for an interrupt while processing, which needs a third state: “listening for barge-in, not writing this turn.” Phone IVR often uses a comfort tone as processing. Drawing processing as still recording also lies on the privacy layer.

Applying it

• The moment the endpoint is declared, switch to a look mutually exclusive with listening, and stop the level animation.
• If processing exceeds budget, show progress rather than an endless spinner.
• Accept with injected delay: speech during processing must not merge into this turn’s transcript unless the product explicitly supports append.

Related

• Same group: C7.16.1 While speaking, continuous visual feedback must prove the system is capturing · C7.16.2 Level or waveform motion lets people tell the mic is receiving sound, not that the device has failed · C7.16.4 Without visible feedback, silence could be a hang or a normal wait
• Adjacent: C7.09 Endpoint Detection and End-of-Utterance · C7.10 Live Partial Results
• Search: processing state · barge-in · listening versus thinking