Auto subtitle generator

Subtitles extend reach in three specific ways. Captions make video accessible to viewers who are deaf or hard of hearing. On social media, the majority of video views happen with the phone volume off: captions keep viewers engaged even in silent-viewing contexts. For multilingual audiences, subtitles allow non-native speakers to follow content more easily.

Generating subtitles manually by transcribing and timestamping audio is one of the most time-consuming parts of video production: a 10-minute video typically takes 45 to 60 minutes to caption manually at professional accuracy. AI transcription compresses that to 2 to 5 minutes of compute time.

The tool runs a two-stage pipeline. In the first stage, FFmpeg extracts the audio stream from the video: it strips the video track (-vn), selects only the first audio stream (-map a:0), downsamples to 16 kHz mono (-ar 16000 -ac 1), and outputs raw PCM float32 data (-c:a pcm_f32le) capped at 30 minutes (-t 1800) to prevent browser memory exhaustion. The resulting WAV is decoded by the browser's AudioContext into a Float32Array.

In the second stage, the Float32Array is passed to Whisper running in a Web Worker via the @xenova/transformers library, which runs entirely in the browser using WebAssembly. The worker returns timestamped segments for editing. If you choose to burn the subtitles into the video, FFmpeg's libass subtitle renderer runs a second pass: Roboto Bold at FontSize 22, white text, 2.5px outline, 1px shadow, and 25px bottom margin.

The Whisper model is downloaded on first use and cached in the browser afterward. The first run is slow; subsequent runs start transcription immediately.

Transcription accuracy depends heavily on audio quality. Clean, close-mic audio with minimal background noise produces near-perfect results. Distant microphones, heavy reverb, or multiple overlapping speakers reduce accuracy. For interviews with two speakers, the output will contain both speakers' words correctly but will not attribute them separately.

Review the generated segments before burning. The editor lets you correct timing and wording before the burn pass, which saves time compared to correcting burned-in subtitles that must be re-exported.

Cloud-based transcription services (Otter.ai, Rev, Descript) upload your audio to external servers for processing. For video containing confidential meetings, legal proceedings, medical consultations, or unreleased content, uploading to a third party may not be acceptable. This tool runs Whisper locally in the browser: the audio never leaves your device. Local Whisper is also free with no per-minute billing.