Designing a Voice
Kokoro reads from a fixed catalogue and Coqui clones from a sample — both work, but neither handles a twenty-plus-character book without a few voices starting to repeat. Qwen VoiceDesign takes a different approach: describe a persona, and it designs a bespoke voice to match, built from how that character actually speaks in your book rather than picked off a shelf.
Open a cast member's profile drawer — a character with no voice yet is flagged "Needs voice" — and Castwright drafts a starting voice persona straight from the text: age, accent, the vocal qualities the writer kept gesturing at. Below, Insp. Cray — carried across from The Drowning Bell into Saltgrave at 97% match confidence — shows the same drawer mid-series: engine set to Qwen (bespoke), a persona textarea with Regenerate for a fresh draft, and a Play 12s sample button to audition the current voice before touching anything.
Edit the persona text (or click "Regenerate" for a fresh draft), pick Qwen
(bespoke) as the voice engine, and click Design & preview. Generating
persona text needs a GEMINI_API_KEY set from Account → Server
Configuration (or in server/.env for CI / power users) — the drawer warns
inline when it's missing.
Design runs on the GPU, one character at a time. The top bar shows a "Designing · N%" pill while it works, and it's safe to close the drawer — progress keeps going in the background, so you can move on to the next character while this one finishes.
A single flat voice gets you who's speaking; emotion variants get you how. Once a character has a designed base voice, the profile samples directly from that design (no library voice matched yet) and the button becomes Design & compare for another pass. Saving pins the voice across the whole series, so it travels with this character into book two the way it should. From here, per-emotion variants — Whisper, Angry, and any other tags used in the manuscript — become available to design individually, so a character who's furious in chapter three actually sounds furious. Below, Wren — a thirteen-year-old apprentice, designed from 13 lines of dialogue — already carries 4 emotion variants.
Doing this one character at a time works, but a twenty-voice cast makes for twenty errands. "Design full cast" (next to the cast table) opens a scope picker instead: Base voices for whoever still needs one, Emotion variants for tagged emotions missing a take, or Both — bases first, then their variants — with live counts of exactly how much work each choice queues. Designs still run one at a time on the GPU under the hood; the picker itself is safe to close while the queue works through it.
Screenshots of the design-in-progress state and the full-cast scope picker are tracked as a follow-up.
Next: Voice Engines.