Final transcription after a call

When a real voice call closes, BeAI queues a final transcription in the background. The selected model receives only the caller audio track and the existing call summary as context. Live text remains available while processing is pending.

After success, caller text is replaced in the conversation, API detail and text export. Bot replies, their dates and their original order are preserved. The call summary and completed actions are not rewritten.

Select a model

Platform administrators open Platform settings → Final call transcription and select OpenAI GPT Transcribe (the default) or Gemini 3.5 Transcribe (Preview). Saving applies the choice to newly closed calls. Jobs already queued retain their selected model. Organisation administrators cannot change this platform setting.

The equivalent API operation is PATCH /api/v1/system-settings/final-transcription:

{"model": "gpt-transcribe"}

GET /api/v1/system-settings returns the value in data.final_transcription. Both interfaces enforce the same platform permissions and allowed models.

Retained and temporary recordings

All real voice calls are captured for this processing. When bot recording is enabled at call start, the original recording is retained and remains available to authorised users as before.

Otherwise, audio stays in private temporary storage, with no player, download or audio API access. It is deleted after successful transcription. When the bot's auto-improvement also requires the recording, transcription waits for that audio analysis to finish. Temporary files are retained for at most 24 hours if processing fails persistently.

The bot recording option therefore controls retention and playback. Simulations, written WhatsApp messages and voice notes do not use this queue. Historical calls are not automatically reprocessed.

Processing states

The conversation detail shows Queued, Processing, Retry scheduled, Completed, or Failed. Reload the page to refresh it. The API exposes the same state in the conversation's final_transcription field. A failed job keeps the live transcript and displays a diagnostic code without sensitive content.

Temporary failures have up to five attempts with increasing delays. An interrupted worker resumes its durable queue. Saved results for completed audio chunks are reused. Text exports always contain the currently visible transcript.

Reading limits

Caller text is placed between bot replies by matching it to historical live turns. Placement remains approximate when the live transcript was inaccurate. Returned words are not manually corrected. Historical times are message log times rather than precise speech boundaries. Gemini provides word timestamps; GPT Transcribe does not provide native timestamps.

The existing summary can contain errors and influence name recognition. Neither model guarantees a completely accurate transcript.