Persist accumulated verbose_json on the recording and broadcast only transcript/whisper deltas so the show page can render language, segments, word timestamps, and confidence while a run is in progress.
Consume faster-whisper SSE instead of blocking on one JSON response, persist growing text and progress over Reverb, and drop the fake percent heartbeat.
Ship FrankenPHP Compose services with Reverb WebSockets for real-time transcription status, skip re-uploading identical audio via content hash, and format disk usage without requiring intl.
Raise queue retry_after above the job timeout, recover orphaned runs, allow stop/restart with any engine, and keep finished transcripts searchable in the recordings list.
ben