openai calls Whisper API (cloud, fast, good on speech). faster-whisper runs Whisper locally (no API call, slower on CPU).

Extracts the vocal stem from the mix before transcription. Strongly recommended for music — without it Whisper hallucinates random languages on instrumental-heavy tracks. First use downloads ~3 GB of model weights (one-time).

htdemucs (v4 hybrid transformer) is the speed/quality default.

When off (default), the vintage toggle is hidden in the UI and any vintage=true sent via the API is silently ignored. Enable only on instances with a local faster-whisper + Demucs stack powerful enough to run it.

Forces local large-v3 + vocal separation regardless of the provider. ~6× realtime on CPU — best run as a one-off via --vintage on the CLI for individual jobs, not as a default.

When the first pass returns a degraded transcript (Whisper covered < 30 % of the audio — the textbook hallucination on instrumental-heavy music), automatically re-run the same audio with Demucs forced on. The user pays credits exactly once regardless of the retry. Off → leaves the user with a degraded transcript and a warning banner instead. Adds wall-clock time only when the first pass actually fell over.

When a job is submitted with Audio type = Music track, automatically run Demucs vocal separation even if the global "Vocal separation" toggle is off. Recommended on — guitar and instrument hallucinations are the #1 cause of poor music transcription. Only disable if Demucs is not installed or on very resource-limited servers.

When a music job is processed, override the configured transcription model to whisper-1 regardless of the global model setting. Recommended on — whisper-1 is the only OpenAI model that returns per-segment timestamps. Without timestamps, LRC and SRT files cannot be time-synced. Disable only if you intentionally want gpt-4o-transcribe on music jobs.

Fraction of audio that must be covered by transcript segments before the result is considered acceptable. Below this ratio the auto-retry safety net fires. Default 0.30 (30 %) is speech-optimised. Raise to 0.50 on music-first deployments to catch more hallucinations (complementary to the blip-pattern detector which works regardless of this value).

Optional. Admin key with api.usage.read scope — enables live billing data on the dashboard. Create one at platform.openai.com/api-keys.

whisper-1 is whisper-large-v2 under the hood — fine for clean speech but hallucinates random languages on music with strong instrumentals (the "Take On Me as Russian" failure mode). gpt-4o-mini-transcribe is the cost/quality sweet spot. gpt-4o-transcribe is best on accents and difficult audio. Pricing changes — check openai.com/api/pricing.

Optional secondary model used if the primary fails (transient API errors, model outage). Pick a different model from the primary so a model-specific outage doesn't take both down.

OpenAI Whisper hard limit. Do not raise above 25 unless the API limit changes.

OpenAI Whisper is $0.006/min. Update here if pricing changes.

128 kbps mono is sufficient for Whisper.

Empty disables real email — messages are only logged. Save applies immediately.

Leave blank to keep the current value (currently: ).

Used to build verify / reset links inside emails. Production must be the public HTTPS URL.