Auto Subtitle Generator

The engine is Whisper, the open speech-recognition model whose tiny and base variants are small enough to run on a laptop CPU. It was trained on 680,000 hours of labelled audio, which is why it handles accents, background noise and domain jargon far better than the phoneme-based recognisers of a decade ago. This page runs the model with onnxruntime WebAssembly, entirely in your tab.

Where the audio comes from

Your video file is decoded with the Web Audio API and only its audio track is kept, resampled to exactly 16 kHz mono — the format Whisper was trained on. The video pictures are never processed; they stay on screen purely as a reference while you review the subtitles. MP4 (H.264), WebM, MOV and MKV all work as long as your browser can play them.

Chunking long videos

Whisper consumes 30-second windows, so a long video is sliced into overlapping 30-second chunks with 5 seconds of overlap. Each chunk is transcribed with timestamps, then the overlap region is assigned to the earlier chunk — words at the boundary are kept once, never doubled. The progress bar tracks this chunk by chunk, so a 15-minute video visibly advances instead of hanging.

Editing before export

Every segment is editable: the text, the start time and the end time. Click a timestamp to jump the video player straight to that line and check it against what you hear. SRT and VTT files are generated from your edited segments, not the raw model output, so what you download is exactly what you approved.

What it cannot do

The tiny model (~40 MB) is fast — roughly real-time on a modern laptop — but it stumbles on heavy accents, crosstalk and technical vocabulary; the base model (~80 MB) is noticeably more accurate at the cost of a longer download and slower transcription. Neither diarises speakers: overlapping voices come out as one stream, and the page does not claim to tell your speakers apart.