Script for an AI avatar: how many words fit in 60 seconds

A script for an AI avatar is the exact text the avatar will say, so write it to be spoken and cut it to the tool's length limit. On Sume, one talking video takes a script that Sume estimates at 4 to 60 seconds, and its current code counts 2.8 words per second: in Sume's own tests, 165 words in one unbroken run estimate at exactly 60 seconds, and 169 words estimate at 64 and are refused.

The 4–60 second window comes from Generate avatar video, read on 2026-09-27. The word rate, the counting rules, and the English-only rule are read from Sume's current code, so they describe today's behavior rather than a documented promise. For material past one minute, see how to make an AI avatar video longer than 60 seconds.

How many words fit in a 60-second AI avatar video?

About 165, but not as a fixed number. At 2.8 words per second, 60 seconds is 168 words; Sume, however, does not divide the whole script at once. Today's code cuts the script into clips, estimates each clip at its word count divided by 2.8, rounds that up to a whole second, counts every clip as at least 4 and at most 12 seconds, and adds the clips together. The request is accepted when the total lands at 4 to 60 seconds, inclusive.

Sume's own tests show how the total moves:

Why isn't the limit an exact word count?

Because the rounding happens clip by clip, and the clips depend on the sentences as well as the length. The code splits sentences at ., !, and ? (and the full-width 。, ！, ？) and words at spaces and line breaks. It groups short sentences into one clip and cuts a long run of words into balanced parts. Two scripts of nearly the same length can therefore estimate quite differently: the 170-word test script, written as ten sentences, estimates at 70 seconds, while 169 words in one unbroken run estimate at 64.

• Near the limit, a word count alone won't tell you whether the script fits. Send it, and trim it if Sume refuses it.
• A multi-scene request shares the same 4–60 second window across all its scenes, and in current code a silent scene counts for at least its duration. A UGC ad script for AI avatar videos budgets a hook, a silent beat, and a call to action inside it.

What happens when the script is too long?

Sume refuses the request with 400 invalid_request. In current code, the error's details.errors entry for script states the estimate: "Avatar video script is estimated at 64 seconds; maximum is 60 seconds. Shorten the script or split it into multiple videos." The docs give the same two choices, shorten the script or split it into multiple jobs; how to make an AI avatar video longer than 60 seconds shows how to join the parts afterwards.

In current code there is no practical minimum: an empty script is refused, but a single word already estimates at 4 seconds, the bottom of the window.

What language and wording should the script use?

English, in current code: the prompt behind each clip says "Spoken language: English only." and asks for "only the exact English dialogue", with nothing translated, paraphrased, improvised, or added. That shapes how you write:

• Write only the words to be spoken. Everything in script goes to the video model as dialogue, so a note such as "(smiles)" counts as a word and becomes part of the line.
• Put the setting in the request's scene prompt or photo, not in the script.
• For speech in another language, make the audio with TTS 1.0, whose language field the Sume API reference documents, and lip-sync a still to it with VEED Fabric 1.0, as in the lip sync API guide.

Where does the script go in the request?

Send it as script on POST /v1/avatar-1.0/talking-video, next to a ready avatar's avatar_handle. The request takes exactly one of script or video_inputs; the scene-by-scene version is covered in the multi-scene avatar video API, and the quality, aspect ratio, and scene options in Talking avatar video API.