Skip to main content
A clip is a talking-avatar video you generate from a script. You write what the avatar should say, generate the resulting TTS audio, optionally edit pronunciation and avatar video instructions, then render the video. This is separate from live conversations. A live conversation is a real-time call, where you talk to the avatar, and we then transcribe your speech, think of a textual response via an llm, then return a spoken response with the resulting avatar video stream. A clip is offline: you supply the inputs (what to say, how to pronounce it, instructions for the avatar during the video, etc..), and Akapulu Labs generates the audio and video.
Today clips are created over the API. Creating and viewing clips in the Akapulu Labs UI is coming soon.Live conversation avatars will also be updated soon so they look like clip avatars.

​
Clips and takes

A clip is a named project. A take is one generation of that script.
Use a clip to group related takes of the same piece, for example several pronunciations or looks of one line. The first take on a clip is marked as the favorite. You can change favorite_id later. To start a take, enter in the text you would like your avatar to say, as well as which avatar you would like to use, and we will generate the spoken audio. From there, you can optionally edit the phonemes (how the line is pronounced) and the video instructions (how the avatar should look and move, i.e. face, emotion, and body). After audio is ready, we split it into chunks of about 2.5–3 seconds and write two prompts per chunk: face (head / expression) and upper body (shoulders, arms, hands). We generate those automatically from the line; you can edit them before you render.
Then render the video. After render, that take is frozen: you cannot change phonemes, prompts, or render it again. Start a new take (from scratch, or by copying an existing take configuration) if you want another version.

​
Why audio and video are separate

Video takes a long time to generate and uses a lot of credits. Audio is relatively quick and cheap, so we recommend playing with the spoken line, i.e. pronunciation, timing, and video instructions, until you like it, then triggering the video job. That is why we have chosen to disentangle audio & video generation, as opposed to one sweep from text to video. Splitting the methods also lets you edit face and upper-body prompts on each chunk before you spend the video credits. Here’s the approximate time and credit cost for a given length of spoken audio/video (generation time is how long the job runs, not how long the clip lasts): Audio Video

​
Quickstart

1. Create a clip — POST /api/clips/create/ with { "name": "Product intro" }. You get back a clip id. 2. Generate audio — POST /api/clips/<clip_id>/takes/ with avatar_id and text (the line to speak). Use a catalog avatar avatar_id, or one of your own (private avatars require an enterprise plan). You get back a take id, status, and the converted phonemes. audio_url is null at first. 3. Poll — GET /api/clips/<clip_id>/takes/<take_id>/ until status is ready_to_render. Then audio_url will be set and you can download the audio. 4. Render — POST /api/clips/<clip_id>/takes/<take_id>/render/ starts video generation. Poll the same GET until status is complete. Then video_url will be set. Video takes longer and uses more credits than audio. For phonemes, video instructions, status values, and credits, see Workflow. Also: Phonemes · AI content disclosure · Python example