Automatic subtitling used to be two separate systems glued together: a speech recogniser that produced words, and a second "forced alignment" step that worked out when each word was said. Modern speech-to-text models do both at once, returning timed segments directly. SubtitleWave sends your clip's audio to one of these models through Vercel AI Gateway and builds subtitles from the timed text it returns.
From sound to subtitles, in five steps
1. Sound becomes a picture. The audio is turned into a spectrogram: frequency content measured many times a second. Speech shows up as bright bands, silence as dark.
For a different kind of learning, ahaboo walks you through subjects like why the Moon shows phases and how photosynthesis turns light into food.
What gets sent, and when
Nothing leaves your device until you press "Generate subtitles". At that point the audio track of your file (re-encoded as a small WAV, capped at 10 minutes and about 25 MB) is uploaded once, transcribed by the model, and the resulting text and timestamps are sent back. The original video file is never uploaded; only its extracted audio is, and only for transcription.
Where the timestamps come from
The transcription model returns a list of segments, each with a start time, an end time and the words spoken in that span. That is why timing is segment-level: accurate at the start and end of a phrase, estimated in between. SubtitleWave splits long segments into readable subtitles (at most 2 lines of 42 characters, about 6 seconds) and shares the segment's time across the parts by length.
Why the first run takes a moment
Every run uploads audio and waits for the model to respond; there is no local model to download and cache, so timing is mostly upload speed plus the model's processing time, roughly proportional to clip length.
Where the AI still needs you
- Names and jargon: the model writes the most likely spelling it has seen, not your brand's.
- Overlapping speakers: the transcript is one stream of text and does not label who is speaking.
- Music and silence: it can invent filler phrases. Scan those sections on the timeline.
- Reading speed: people speak faster than viewers read. The SRT editor flags lines over 17 characters per second so you can shorten them.
Language models on top
Transcription is one Gateway model; some jobs use a second, text-only one: translating a whole subtitle file naturally, fixing punctuation, condensing lines that are too fast, or proposing YouTube chapters. For those, SubtitleWave sends only the subtitle text (never the audio) to a language model through the same Gateway and puts the results back into the same timings. Try it in the subtitle translator.