Parameters 6 min read
Auto-Subtitle Accuracy: What Hurts It and How to Improve It
Local speech recognition (Whisper-class models) reaches high accuracy on clean single-speaker audio and degrades predictably: noise, overlapping speakers, heavy accents and domain jargon are the four killers. All of them are addressable before transcription.
Pre-transcription checklist
- Denoise if there is audible background hiss/hum — but lightly; over-denoising smears consonants.
- Normalize loudness to around -16 LUFS so quiet passages do not fall below the model floor.
- Extract mono 16 kHz audio (what the model consumes anyway) — long videos transcribe faster as audio-only.
- Specify the language explicitly instead of auto-detect when you know it; fewer misfires at clip starts.
Post-transcription pass
Expect to hand-fix: names, technical terms, numbers with units, and sentence boundaries in long pauses. Editing a generated SRT in a timeline editor is faster than typing from scratch — even 90%-accurate output beats a blank page.