Persistent HTTP 500 errors on whisper-large-v3-turbo audio transcription endpoint

Forum
1

I am experiencing consistent HTTP 500 Internal Server Error responses from the whisper-large-v3-turbo transcription endpoint (POST /openai/v1/audio/transcriptions), starting from April 10, 2026.

Audio files are split into ~5-minute chunks (~500KB–2MB each, Opus format, 16kbps mono)

  • The first chunk consistently succeeds (HTTP 200)

  • All subsequent chunks immediately return HTTP 500, regardless of their content or size

  • Retries on the same chunk always return 500 — the error is deterministic, not transient

  • We have verified we are well within rate limits (used ~3,900 audio seconds out of 4,000,000 daily limit)

  • The issue persists across different audio files and different days

  • Your status page shows no active incidents

What I have tried:

  • Re-encoding each chunk independently (instead of stream copy) to ensure valid headers

  • Reducing chunk size (from 10 minutes to 5 minutes)

  • Disabling SDK-level retries