Use 2–12 seconds of clean, single-speaker speech. Conversion changes the voice in a recording; expression creates English speech from text. Listen before using the result.

English. Use [laugh], [chuckle], [sigh], [gasp] or [cough] in the words. A target voice needs 5.1–12 seconds of clean speech. Event delivery needs listening review; expression/pacing sliders apply to the original model.

Audio/transcription: up to 60 seconds. Video uploads: up to 100 MB / 1080p.

Edit the words and timing below. The original script or transcript stays saved. Exports keep up to 30 seconds and 0.5 megapixels. A new soundtrack does not change mouth movement.

Turn this off for narration only. Subtitle files are still saved.

The larger model can take longer. Both need word review; neither identifies overlapping speakers.

English ↔ Spanish translates locally. Correct the transcript text first; its original is kept. Review names and meaning. Enter all target lines below to use your own translation; leave every line empty for automatic translation.

Dubbed video exports: up to 30 seconds. Lip-sync previews: up to 8 seconds. Dubbing replaces the original soundtrack; music and background sounds are not separated.

Compare methods on the same face and speech. LatentSync allows 0.5–4 seconds, needs 18 GiB free GPU memory and uses adapted face alignment. Profiles and hidden mouths remain unsupported.

Optional: mark when a hand or object covers the mouth. The entire original frame is kept in those ranges, with a short blend at each edge. Audio continues. This does not detect occlusions or repair side views.

0.5–8 seconds of speech. One clear, nearly frontal face. The mouth area is generated; lip and tooth quality need playback review.

Installed preset voices; no voice cloning. Dubbing keeps pauses and uses bounded pitch-preserving timing adjustment.