Troubleshooting¶
Resolve Whisper CPU startup, audio input, timestamp, and TLS problems.
The API service is not active¶
Fix:
sudo cloud-init status --wait
sudo -iu ubuntu
sudo systemctl is-active sushi-whisper-cpu-api.service
sudo journalctl -u sushi-whisper-cpu-api.service --no-pager -n 50
Do not recreate the instance identity to force a retry.
The HTTPS call rejects the audio¶
Cause: the clip is over 10 MiB or five minutes, the format is not WAV, FLAC, or MP3, or another request is still running.
Fix: stay inside those limits for sushi-whisper-cpu-api, one request
at a time. For a larger file on version 0.1.10, use the local SDK in
Configuration. Versions 0.1.8 and 0.1.9 cannot
use the direct-filename SDK call.
Timestamps look wrong¶
Cause: the API returns text, segments, and timestamps. Accuracy depends
on the recording and the instance. The documented check is to listen to the
original and review transcript.json.
Fix: compare names, numbers, and timestamps with the source audio before you keep the transcript.
The workstation call fails TLS¶
Fix: forward port 8443 and trust server.crt from this instance. On the
instance, sushi-whisper-cpu-api reads the account file.
Last updated November 1, 2026 · Suggest an improvement