Custom avatar with CUSTOM voice renders the wrong face (stock-like fallback) in gwm1_avatars realtime sessions; preset-voice clone of same reference renders correctly

Summary

A custom Character configured with a custom voice consistently renders a different face than its stored reference/processed image in gwm1_avatars realtime sessions. An otherwise-identical Character (same reference image) using a preset voice renders the correct face. This looks like the custom-voice configuration routing through a path that fails to load the custom avatar and falls back to a stock-like face.

Reproduction (ours, repeated across ~10 sessions in one day)

1. Character ddc7d594-b940-42e4-8754-d1e4d37fca13 (name COLTRANE, custom voice id 10fd1e64-dbc7-43da-8414-b77b70889342, status READY). Its referenceImageUri and processedImageUri (verified via GET /v1/avatars/:id) show the intended face.
2. Create a realtime session: POST /v1/realtime_sessions with {model: 'gwm1_avatars', avatar: {type:'custom', avatarId:'ddc7d594-…'}, livekit:{…}} (puppet mode — we publish audio, subscribe the avatar A/V).
3. The published avatar video is a different person than the stored reference — every session. Example session ids from 2026-07-02 (US MST): 86336814-09eb-47c6-8309-0ad404e28fb8, dc79d811-… era runs, da69b3b0-022b-4184-8809-2435ada5ac1d, dabad714-eadb-4d00-a996-d5ce1fe829b4.

Discriminating experiment

Created a5bf9339-2f48-4c60-aca9-78205a39a89a (COLTRANE-PRESET-AB) from the same reference image, with preset voice jasper — session dc79d811-d0a0-42d6-a066-8937dc9e92de rendered the correct face immediately. Same night, same integration, same room. Only variable: voice type.

Possibly related: docs note webcam/screen-share is unsupported for custom Characters with custom voices — the same configuration that fails here.

Environment

• API version header 2024-11-06, raw REST integration (also reproduced via @livekit/agents-plugin-runway path)
• Happy to provide org id, more session ids, or run a live repro.

Expected

Custom avatar + custom voice renders the stored processed image, or the session errors loudly instead of silently rendering a different identity — for avatars representing real/branded personas, silently wearing another face is a serious failure mode.

Start by reproducing POST /v1/realtime_sessions with the reported custom avatar and custom voice, then compare it with the same reference image using preset voice jasper. Check the @livekit/agents-plugin-runway path and the custom-character webcam/screen-share limitation. Done means the custom voice session renders the stored processed image or fails explicitly instead of showing another face.