Video to Subtitles: The Local Speech-Recognition Pipeline
Lecture recordings need subtitles, interview footage needs transcripts, vlogs need a silent-scrollable version — all start from the same step: generating text from the speech inside a video. Browser-local recognition (Whisper) means not one byte leaves the device. The pipeline runs in three stages: preparation, generation, proofreading.
Two output routes
The subtitle route: timed subtitle files (SRT/VTT) whose lines map to the picture — for attaching to videos or uploading to platforms. The transcript route: plain text for notes, archives and re-writing. Same recognition engine, two shapes — and the proofreading emphasis differs (subtitles live or die on line breaks; transcripts on readability).
Two tools map to the two routes: auto-subtitle (SRT/VTT out) and video-to-text (plain text out), both local, both taking video files directly (the engine extracts the audio internally).
Preparation: audio quality decides everything
Recognition accuracy’s number-one variable is the audio: clear speech at moderate volume with low noise floor reaches over ninety percent; noisy rooms, overlapping voices and music burying speech fall off a cliff. Before recognizing, check the source: denoise where possible (denoising costs a little fidelity, but recognition gains outweigh it) and lift quiet tracks to a sane loudness.
Video itself is no obstacle — the engine only hears the track. But mixed content where speech competes with music (MV-style vlogs) is naturally weaker than talking-head material; set expectations accordingly.
Generation: timing and process expectations
Local speed tracks device performance: an hour of audio typically takes several to a dozen-plus minutes on a modern laptop, all inside the browser with nothing uploaded. Long videos should be split by chapter first — single runs beyond two hours raise memory pressure and make retries expensive.
Generation is unattended: start it, wait for the progress bar, download the subtitle file.
Proofreading: proper nouns and line breaks
Three mandatory checks: proper nouns (names, brands, terminology — the engine’s weakest area, replaced one by one), punctuation and line breaks (automatic segmentation sometimes cuts at odd points; adjust by meaning), and numbers with units (“thirty minutes” versus “30 minutes” consistency). Transcripts add readability cleanup and verbal-tic trimming.
The effort-saving rule: subtitles get stricter proofreading than transcripts — subtitles are read directly by viewers, transcripts only need semantic readability.
Format choice and how to attach
Formats: SRT is the universal default (players, platforms and editors all read it), VTT targets web players, ASS carries styling (fonts, positions, effects) — the full comparison lives in the subtitle-format guide. When unsure, SRT.
Attachment splits two ways: soft subtitles (a separate file loaded by the player, switchable) and hard subtitles (burned into the picture, visible everywhere) — the trade-offs live in the subtitle-methods guide. Platform publishing usually uploads the SRT and lets the platform attach it.
The privacy advantage of local recognition
Interview footage, internal training, medical and legal content — speech like this is sensitive, and cloud recognition uploads it to a third party. Local recognition (Whisper running in the browser) keeps every byte on the device, a structural advantage for privacy-sensitive material. With the pipeline built, subtitle cost drops from per-minute outsourcing rates to local electricity.