Lip sync a Korean voice recording on Sume: which route, what it costs

For a Korean recording, skip the script-driven Avatar Video route and use an audio-driven one: MiniMax H3 Max lip-sync (5 to 14.8 second audio, $0.10 per second at 768p) or VEED Fabric 1.0 ($0.1875 per second at 720p). Both take a still image or ready avatar plus Sume-hosted audio and make the mouth follow it. The talking-video route's clip prompt is English only, so it is the wrong tool.

Preparing the recording

Both routes need audio Sume can fetch, which in practice means a Sume-hosted URL. If your Korean voice is inside a video, extract it first with audio detach, which writes a wav by default. The timeline audio page notes that wav is the format to keep when a file drives lip sync, because mp3 adds priming padding at every edge. Audio detach is a flat $0.01 per job, per the OpenAPI text for that route.

Cut the audio to the window. For H3 Max that is 5 to 14.8 seconds, so a 40 second recording becomes three or four segments.

Cost of a 12 second line

For a single 12 second Korean line:

• H3 Max at 768p: 12 x $0.10 = $1.20.
• H3 Max at 1080p: 12 x $0.20 = $2.40.
• Fabric at 720p: 12 x $0.1875 = $2.25.
• Fabric at 480p: 12 x $0.10 = $1.20.

What Sume does not claim

The route descriptions say the mouth follows the audio. They do not publish a language list or a lip-sync accuracy figure, so the Korean result is something you verify. Make one 5 second test at H3 Max 480p, which is 5 x $0.0625 = $0.3125, and watch the mouth against the audio with the sound on. If it passes, scale up. Add Hangul subtitles with the korean-ad or another Hangul style through Video captions.

Which route to pick

Use H3 Max when the audio fits 5 to 14.8 seconds and you want the lower rate. Use Fabric when the recording is longer than that window, since its priced range runs up to 300 seconds, but note it costs $0.1875 per second at 720p. Do not use the experimental sume/avatar-1.0/fabric route in production: its description says it is test-only and that the model id will change.

Whichever you choose, send exactly one visual source: either an image_url of a posed still or an avatar_handle, never both.