Retrain a cloned voice without breaking old videos

On Sume, an avatar is referenced by its avatar_handle, so videos you already made keep working and new requests keep using the same handle. For voice continuity, read the voice, model_id and settings that each text-to-speech job records, and reuse them for the next line. The docs read here describe no retrain operation.

What HeyGen's release says

The HeyGen September 2026 release page says retraining a professional voice clone preserves the voice ID, name, language and slot. If retraining fails, it falls back to the previous voice, and deleting a voice frees the slot. The practical point is a stable identifier for downstream work.

Sume's stable identifier: the handle

POST /v1/avatar-1.0/generate creates an avatar under a top-level avatar_handle, stored without a leading @. Later POST /v1/avatar-1.0/talking-video requests refer to the ready avatar by that handle. The Avatar docs describe creating from a prompt, structured traits or a photo; they do not describe retraining an existing avatar, so a changed identity means a new avatar under a new handle.

Keeping a voice consistent

A completed text-to-speech job (text_to_speech) records how its audio was made: model_id, voice ({"mode": "id", "id": "..."}), language, output_format, and the settings it used, generation_config and speed. Each setting is null when the request did not send it. The jobs docs say to read these from the job to make the next line sound the same.

A change plan that does not touch old videos

Completed jobs and their media.sume.com artifacts stay readable by job id, so a new avatar or voice never rewrites a finished video. When you change identity, give it a new handle or version suffix such as presenter_v2, run a short test video with the new handle, and move your pipeline over only after you approve the result. Keep the old handle until nothing references it.

A job is tied to its workspace and creating member, so record which key made it. See Jobs and results.