Whisper Repetition Errors: Beyond the 30-Second Sample
The problem appeared after the sample
I asked AI to transcribe a Korean recording with a large Whisper model and keep downloads and caches on an SSD. It tested a short section before processing the full recording.
The reported timings for the 30-second sample were:
| Execution method | Model | Reported time |
|---|---|---|
| openai-whisper with MPS | large-v3 | About 46 seconds |
| MLX-based execution | large-v3 | About 7 seconds |
AI used the MLX approach for the full recording. The output then repeated the same short phrase through part of the transcript. A file had been produced, but the result was not ready to use. The short sample had not exposed the problem in the longer output.
These are two historical execution reports, not a controlled benchmark with repeated measurements. They do not establish the same performance difference on other recordings or devices.
A settings change, followed by specific checks
After detecting the repetition, AI reran the recording with the same large-v3 model. The recorded change was condition_on_previous_text=False.
The useful part of the report was what happened after that change. AI checked for the repeated pattern and compared the final segment’s end time with the end of the recording. It reported replacing the faulty first output with the new transcript.
| Check | What the report established |
|---|---|
| Repeated-pattern check | The pattern being checked was no longer detected |
| Final segment end time | The output extended to the end of the recording |
| Sentence-by-sentence comparison with the audio | No evidence of a complete comparison |
A transcript can pass a repetition check and still contain misheard words. A matching end time does not rule out missing speech in the middle. These checks addressed the observed failure; they did not measure overall transcription accuracy.
Carrying the checks into the next file
The record then shows another recording processed with the same settings and an end-of-audio check. The revised procedure was used on a subsequent input.
I specified the recording, model size, and storage location. AI compared execution methods, detected the repetition, changed the setting, and reran the job. The completion report provided a repetition check and the final segment time, beyond confirmation that a file existed. It did not establish sentence-by-sentence agreement with the audio.
This account comes from work records dated May 26, 2026. The timings and checks were reported then and were not repeated for this article. The recording’s contents and original filename are omitted. The record supports a correction to a particular repetition failure, while full transcript accuracy remains unverified.