Transcription batches

Open Transcription → Bulk import to upload a collection for processing later. These uploads are held before any provider call, even when ordinary uploads automatically transcribe. The normal Upload page keeps its existing policy.

Prepare and run

• Waiting recordings are grouped by the validated date in their original filenames, newest first. Explicit timezone offsets do not change that calendar date. Files without a recognized call timestamp appear under Recording date unavailable.
• Select a date or a smaller selection from it. Save a named batch for another day, or start it immediately. A batch contains up to 1,000 recordings from one date group. Saving reserves those recordings so another batch cannot take them. Only an unstarted batch can be edited or removed; removing it returns its recordings to the waiting list without deleting the files. Started membership is fixed.
• Pause stops additional recordings from starting. An already submitted recording finishes, including related AI processing. Resume requires an explicit action.
• Completing one batch does not start another saved draft. Other batches you already started remain eligible for the shared dispatcher. Batches and history remain available across sign-outs, refreshes, and application restarts.

Each user manages their own recordings. Company provider/model restrictions and existing access and billing checks still apply. A batch uses the provider and model shown when saved; changing deployment defaults does not silently switch it.

Usage and failures

Audio duration is a planning aid, not a provider credit balance or monetary cap. Transcription may also generate a summary, transliteration, and search embedding. Ordinary automatic work elsewhere in the instance can still use provider credits.

The shared MongoDB dispatcher allows one recording pipeline at a time across the instance, including automatic transcription from ordinary uploads, manual requests and cloud jobs awaiting callbacks. Errors stop the affected batch. Uncertain outcomes need explicit review before retrying because the provider may already have charged. Ordinary single-recording requests also ask for acknowledgment before an uncertain retry. Provider callbacks carry an attempt identity to reject stale completions.

Use Retry unfinished recordings after resolving the displayed issue, or Resume enrichment when the transcript is already saved. The retry covers all failed or uncertain recordings in that batch and preserves completed work. Insufficient prepaid credit, operator pause, provider quota and provider configuration have different remedies. Results marked for administrator review cannot be retried from the batch card.

Transcripts and successful AI steps are checkpointed. Retrying failed enrichment uses that transcript and skips already saved steps, without retranscribing audio. Manual results, Soniox callbacks and ElevenLabs callbacks use the same transcript acceptance and charge-once completion path. A saved transcript with unfinished AI appears as Transcript saved, with an enrichment remedy instead of a success notification. Notifications are deduplicated per accepted result and outcome.

New AI plans pack adjacent indexed multilingual text into bounded requests while retaining speaker, word and timing metadata locally. Summary chunks cover the whole recording; intermediate requests state their 192-token limit and the final summary remains limited to 512 tokens. Accepted v1 plans and checkpoints retain their original prompts, models, request identities and chunk layout.

An accepted asynchronous request with no completion after 24 hours is marked for review; it is never resubmitted automatically.

Operations

Keep /api/cron/process-ingestion scheduled every five minutes with the existing cron bearer secret. Starts, resumes, and callbacks also wake the dispatcher, while cron repairs interrupted reservations and job creation and continues approved work. MongoDB stores imports, batches, execution attempts, and the shared dispatch slot; no Redis, extra worker service, or additional dependency is required.

Bulk imports are initiated through the web manager in this version. Scheduled batch starts, provider balance synchronization, and monetary limits are not implemented.