Real-time transcript issues | User Guide

Explain no transcript, delay, missing lines, and interruptions based on recognition drafts and commit behavior.

Transcript mechanism

Recognition chain: the frontend converts audio into 16k / 16bit / mono PCM and pushes it into the recognition stream.
Display mechanism: recognizing writes a draft first, and recognized or the 2-second compensation timer commits it as a formal message.
Message constraint: the transcript list keeps only the latest 40 items.

Frequent symptoms and diagnosis

There is a draft but it does not commit: first treat it as recognition not yet finished or still inside the compensation window.
There is transcript text but no auto answer: first confirm whether it is interviewer text and longer than 3 characters.
Occasional write failure after stopping: /message returning 409 Session already stopped is an end-of-session protection.

Standard troubleshooting steps

Confirm whether the session state is "recording."
Confirm whether both input sources, microphone and system audio, are valid.
Observe whether stable committed text appears within 30 seconds instead of checking only drafts.
If no new text appears continuously, rebuild the session and retest directly.

Acceptance criteria

Drafts refresh continuously and can periodically commit as formal messages.
Key interviewer sentences can trigger auto answer when trigger conditions are met.
No continuous write behavior appears after stopping.

Whether a record is generated after the session ends depends on whether upload conditions are met at stop time. A successful stop does not mean a record is definitely written.

Audio and permissions

Check input sources and sharing authorization first.

Technical issues

Review message writing, answer triggers, and status-code semantics.