← All research
Qweek Research

Roadmap: Proprietary Voice Models

Research in progress on training workspace-scoped voice models for dialogue sync in Motion mode, integrated with Character Bank voice_model_id fields.

Status

Research in progress. This document outlines our intended approach; no production voice training pipeline is live yet.

Goal

Map each Character Bank entry's voice_model_id to a consistent, workspace-owned voice that lip-syncs reliably in MOTION mode — reducing dependency on third-party TTS variance across episodes.

Planned approach

  1. Data — Curated dialogue samples per character with consent and workspace isolation.
  2. Training — Fine-tune or adapter-tune on ElevenLabs-class or open voice models per workspace tier.
  3. Inference — Worker pipeline selects character voice from bible before motion generation.
  4. Evaluation — MOS-style consistency checks across N shots in a sequence.

Non-goals (initial)

  • Real-time voice cloning from single samples
  • Cross-workspace model sharing without explicit export

We will publish metrics and methodology when the pipeline reaches design-partner trials.