Skip to main content
Belcenter AIHelp center
Monitor conversations

Final post-call transcript

Understand transcription modes, follow processing of the caller’s audio and learn how audio is retained.

When a voice call ends, BeAI automatically queues a final transcription for processing. A background service retranscribes only the caller's voice. It also receives the call's existing summary as context.

The live transcript remains available to view during this processing. Upon completion, the user portion is replaced in the history, API details, and text export. The bot's turns remain identical, with their original timestamps and order. The summary and already executed actions are not rewritten.

Processing mode

A platform administrator opens Platform settings → Final call transcription, selects Standard mode or Alternative mode, and saves. Technical service names remain masked. The selection applies to calls closed after it is saved; jobs already in the queue retain their mode. Organisation administrators cannot change this global setting.

The equivalent operation is PATCH /api/v1/system-settings/final-transcription. GET /api/v1/system-settings returns the setting in data.final_transcription. The technical schema is restricted to platform administrators and describes the permitted values.

Retained or temporary recording

All actual voice calls have an audio capture required for this processing, regardless of the bot's recording option.

  • If bot recording is enabled at the start of the call, the audio is retained and remains accessible to authorised users, as before.
  • Otherwise, the audio is placed in a private temporary storage area. No player, download, or API audio access is provided for this file. It is deleted after successful transcription.

If bot auto-improvement also needs to analyse this audio, transcription waits for that analysis to complete before processing and deleting the temporary capture. In the event of persistent errors, temporary file retention is limited to 24 hours.

The bot recording option therefore controls audio retention and playback. Simulations without a real voice session, written WhatsApp messages, and voice notes do not go through this queue. Historical conversations are not automatically reprocessed.

Track the outcome

In conversation details, the Final call transcription section displays:

Status Meaning
Pending The call is in the queue, potentially awaiting audio analysis.
In progress A worker is processing the audio.
Retry scheduled A temporary error will be retried.
Completed The user portion has been replaced.
Failed The live transcript remains available; a code facilitates diagnostics.

Refresh the page to update the status. The API returns the same information in final_transcription on the conversation. Text exports use the currently visible transcript. No temporary files, provider secrets, or request payloads are exposed through this status.

Temporary errors are retried up to five times with increasing intervals. A service interruption leaves the job in the queue to be resumed. Audio segments that have already been processed are reused once their result has been successfully saved.

Usage and cost estimation

Estimates are hidden by default. A platform administrator can temporarily enable Purchase estimates in their avatar menu. The breakdown in the conversation details then displays processes, quantities, unit prices, currencies and subtotals, including the final transcription. The preference is not saved in cookies or the database; closing the browser resets it. Depending on the browser, closing all BeAI tabs or navigating to another page may also reset it.

The estimated provider cost totals the voice conversation, its processing, the final transcription, and any potential auto-improvement. This estimate does not change the amount billed to the customer or their wallet. Every usable response is accounted for, including before a retry; a reused result is not counted twice. A response that could never be recorded may be missing.

Depending on the mode, quantities are audio seconds or tokens; token counts are never fabricated from durations. A missing price or incompatible currencies result in a partial estimate. The purchasing price scale is available to platform administrators after enabling the option, displayed as neutral processes and units. Each response retains the pricing scale locked in at the time of processing.

The detail API GET /api/v1/conversations/{id}?include_purchase_estimates=true exposes the same estimates solely to platform administrators. Without this parameter, estimates remain hidden. estimated_cost already includes the final transcription: do not add its subtotal a second time. Pricing scale configuration remains restricted to the platform.

Reading limitations

User utterances are repositioned between bot messages by matching them with the turns of the live transcript. This positioning remains indicative if the live transcript was incomplete or incorrect. Words returned by the model are not corrected manually.

Historical timestamps are logging times, not precise measurements of speech onset. The accuracy of timestamps depends on the transcription mode; some modes do not provide word-level timestamps. The summary itself may contain errors and influence recognition, particularly of proper nouns. No model guarantees error-free transcription.

Documentation updated with the application