AI Auto Captions for Any Video Online

OpenAI Whisper compiled to WebAssembly via Transformers.js. Four sizes are available: Tiny (~40 MB, fastest), Base (~75 MB, balanced), Small (~240 MB, accurate) and Large v3 Turbo (~600 MB, broadcast quality, hidden behind the Advanced models toggle). Models download once from Hugging Face and are cached in your browser cache for subsequent runs. WebGPU acceleration is used automatically when available.

Yes. Silero VAD pre-segments the audio into actual speech regions before Whisper runs, which both improves accuracy on long files and skips silent intros, outros and pauses. Captions appear in the editor live as Whisper decodes — you do not have to wait for the full transcription to finish before reviewing the start. Large videos (over 30 minutes) work best on a desktop with at least 8 GB of RAM.

You can export to SRT, VTT or ASS — VTT is the standard format for WCAG / ADA / EAA closed captions on web players (HTML5 , YouTube, Vimeo). Captions include timing, speaker labels (when speaker detection is on) and clean punctuation. The burned MP4 places captions in safe zones and supports stroke + shadow for high-contrast readability.

Whisper accepts an initial prompt — text it conditions on before generation. We pass your glossary terms as that prompt, so brand names ("Anthropic"), product names ("Claude", "GPT-4"), proper nouns and technical jargon spell correctly. This is the single biggest accuracy win for tech podcasts, conference talks and brand-heavy content. One term per line, no special formatting.

Whisper infers word boundaries via cross-attention decoding. Accuracy is around ±200 ms — suitable for karaoke-style highlighting and most caption use cases. If decoding is too slow on your device, switch on Fast mode and words distribute linearly inside each segment. Confidence flagging marks words Whisper was uncertain about so you can verify them quickly.

Yes. Opus-MT models handle the top ten language pairs (en, ru, es, fr, de, pt, it, ja, ko, zh). If a direct pair is missing, the tool falls back to NLLB-200 distilled. Keep the original as a second caption line for bilingual output — useful for language learning content and global audiences.

Yes. The tool downloads Twemoji PNGs for every matched keyword and composites them onto the video via FFmpeg overlay filters, positioned above the caption line and timed to the spoken word. SRT and VTT exports also keep emoji characters.

SRT (universal), VTT (web/HTML5 closed captions), and ASS (full animation with all template styles preserved — opens in Aegisub, DaVinci Resolve, mpv or any libass-compatible player). The burned MP4 uses H.264, AAC and the original audio stream copied without re-encoding.

Once a Whisper model is cached after the first run, the tool runs entirely in your browser via WebAssembly — transcription, styling, burning and export all execute locally. Translation models also cache on first use. The page itself loads from our servers, but the actual processing of your video does not require an internet connection.

Yes, in modern Chrome and Safari. On devices with less than 4 GB of RAM, the Small and Large v3 Turbo models are hidden and we default to Tiny to avoid crashes. For long videos a laptop is recommended.