Re-transcribe a Call with Whisper (Fallback Transcription)

When to use this

Your call recordings are automatically transcribed by our voice provider as the call happens. Most of the time that transcript is what you read on the conversation page. Occasionally — for reasons that are usually a quirk of the underlying audio, not the call itself — the automatic transcript drops part of the conversation: a missing turn here, a long silent gap that wasn’t actually silent, a name spelled three different ways.

When that happens, you can run a fallback transcription through a different provider (OpenAI Whisper) on the same audio. It appears alongside the original — the original is never modified — so you can compare and decide which to trust.

This is a stopgap. The primary transcript is still the one created by the voice provider during the call.

How to run it

1. Open the conversation you want to re-transcribe.
2. In the transcript section, look for the Re-transcribe (Whisper) button.
   • If you don’t see the button, the feature isn’t enabled on your account. Reach out to your account contact at TriLuna.
3. Click it. A confirmation pop-up explains the cost and limits — review and accept.
4. Wait. Re-transcription typically takes 30 seconds to two minutes depending on call length.
5. When it finishes, an amber panel appears above the original transcript with the Whisper version. Both stay visible.

What you’ll see

The Whisper panel shows:

• The provider (openai-whisper)
• A timestamp of when the re-transcription completed
• The cost in dollars for this specific run
• Each spoken segment with a timestamp, in order

If the call was over ~24 minutes long, the panel will show a “⚠️ Partial transcript” banner. Whisper can only process about 24 minutes of audio at a time, so longer calls are trimmed to the first 24 minutes before transcription. The rest of the audio still plays normally; only the Whisper transcript stops at 24 minutes. (The original transcript still covers the full call.)

Cost

About $0.006 per minute of audio transcribed:

Re-transcribing the same call twice will cost twice as much — there’s no caching.

What the Whisper transcript is good at

• Catching turns the original transcript missed entirely
• Filling in mid-conversation gaps
• Picking up softer audio that the primary provider’s voice-activity detection skipped

What it isn’t good at

• Speaker labels. Whisper doesn’t know which voice is the agent and which is the caller. The Whisper panel shows what was said and when, but not who said it. The original transcript still has speaker labels — use both together when speaker identity matters.
• Names and unusual words. Like all general-purpose speech recognition, Whisper sometimes guesses wrong on proper names, company names, and domain jargon. Confirm against the original transcript.
• Calls longer than 24 minutes. Only the first 24 minutes are processed (see partial-transcript banner above).

Downloading both transcripts

The Download JSON button on the conversation page includes both transcripts when a Whisper re-transcription exists:

• whisper_rerun — the Whisper version (text + timestamped segments + cost + a truncated flag if the call was over 24 minutes)
• transcript — the original transcript, unchanged

If no Whisper re-transcription has been run for that call, only transcript is included.

Why is this a fallback and not the default?

The primary transcript is created by the same provider that handles the live call, so it benefits from context the audio file alone doesn’t carry (turn boundaries, agent vs. caller separation, real-time adjustments). When it works well — which is most of the time — it’s better than Whisper. Whisper is here for when it doesn’t.

Questions

If you keep seeing incomplete transcripts on calls where Whisper recovers the missing parts cleanly, tell us — that’s the signal we use to push for an upstream fix and decide whether to widen access to this feature.