The page freezes during transcription
Try a smaller or more quantized model, reduce competing tabs, and let the runtime fall back from GPU if necessary. Long recordings magnify memory problems that short test files hide.
Troubleshooting
How to debug browser AI transcription problems
Most issues come from resource mismatch rather than mysterious failure. When the browser, model, and job size are misaligned, symptoms usually show up as freezes, timeouts, or unstable acceleration.
GPU acceleration does not seem to help
Some devices expose acceleration but do not sustain it well for large jobs. In that case a stable CPU or WASM path can produce better end-to-end results.
Folder jobs slow down over time
Queue workloads can expose accumulated pressure. Break very large batches into smaller sessions or use paced processing if the browser begins to degrade.
Export files look incomplete
Confirm that transcription fully completed before exporting. Interrupted or partially failed jobs can leave outputs in a draft state.
The model never finishes loading
Model loading may need a network connection and enough browser storage. Check the connection, available disk space and browser storage, and whether a privacy extension blocks the model host. Save completed transcripts before reloading because unfinished work may be lost.
The transcript contains plausible mistakes
Check that the selected language matches the speech. Listen again to overlapping speech, names and quiet sections. Whisper can produce plausible text where audio is unclear, so verify the result against the source before treating it as a record.
Include these details when asking for help
Report your browser and version, operating system, device memory, selected model, runtime status, recording duration and the last status message. Describe the steps that reproduce the issue. Do not send private recordings; a short sample you are allowed to share is preferable.