Griffin-style listening beats in a recorded avatar video: what works
Tavus Griffin nods and says mm-hm while it listens. A recorded Sume avatar video cannot listen, but it can hold a silent beat. Here is a preview-first test.
Tavus describes Griffin as a full-duplex avatar: it listens and talks at the same time, gives backchannels such as "mm-hm", and stops when you interrupt it. The page says it generates 720p video in 320 millisecond chunks, with 0.43 seconds from audio to video on H100 GPUs, from a single user photo. It also says Griffin-Lite is a research preview for select testers and not available to customers. I read the page on 2026-10-04.
Those are live conversation traits. A recorded Sume avatar video does not listen, since nobody is in the loop. What you can borrow is the pacing.
What a recorded video can and cannot do
| Behaviour | Griffin (per Tavus) | Sume talking video |
|---|---|---|
| Listens while talking | Yes, full-duplex | No, one scripted take |
| Backchannels | "mm-hm" and similar | Not a documented feature |
| Handles interruption | Stops when interrupted | Not applicable to a recorded file |
| Input photo | Single user photo | Prompt, profile or public photo |
| Customer access | Griffin-Lite: select testers only | Public API |
Borrowing the pacing: silence beats
A talking-head video that never pauses feels like a recording. Sume's video_inputs lets you add scenes with voice.type: "silence", which are beats with no speech and a required duration. Use a one or two second beat after a question, as if the avatar were waiting for an answer. The total planned duration must still fall between 4 and 60 seconds, and one avatar and one shared scene serve the whole video.
What the avatar's face does during a silent beat is up to the model. The docs do not promise a nod, so check it on a preview before you pay for a render.
curl -X POST https://api.sume.com/v1/avatar-video-previews \
-H "Authorization: Bearer $SUME_API_KEY" \
-H "Content-Type: application/json" \
-d '{"avatar_handle":"@host","aspect_ratio":"9:16","video_inputs":[{"id":"q","voice":{"type":"text","script":"Ever wonder why your first cut is never the best one?","duration":4}},{"id":"beat","voice":{"type":"silence","duration":2}},{"id":"a","voice":{"type":"text","script":"Because you were still deciding what to say.","duration":4}}]}'Cost of the test
A 10 second video at standard costs 10 x $0.184 = $1.84, and at plus $2.45. The preview comes first, and its still does not change with the tier you choose later.
When you actually need Griffin-style
If the viewer must talk back, you need a real-time avatar product, and Sume does not ship one. Griffin-Lite is not open to customers per the Tavus page. For a one-way message, the silence beat gives a conversational rhythm at a known per-second price.
Sources
Related posts
More in Sume Avatar 1.0
- Avatar validation failed: fix it, or skip verification
Searches for HeyGen avatar validation failed show people stuck on verification. On Sume an avatar create is a job you poll, and failures return a job error.
- Book launch avatar video: use the cover as the product image, cost
An author avatar introduces a book with the cover as the product image. A 30-second clip costs $5.82 at standard, $7.74 at plus and $17.40 at max.
- Car dealership vehicle intro: avatar video with the car as product
Use a car photo as the product image in an AI avatar video. A 20-second intro costs $3.88 at standard, $5.16 at plus and $11.60 at max. See the request.
- Crowdfunding intro video: preview first, render at Max
Check a 30-second campaign intro on a Sume preview, then override quality to max only at generate-video. A 30-second Max render costs $16.50.
Written by Sume