Synthorea / SynthGPT_
A natural-language to u-he Diva synthesizer-preset generator. A prompt like "warm analog bass with a soft attack" is transcribed into a synth-agnostic semantic intent, mapped to an 881-dimensional Diva parameter vector by a family-conditioned mixture-of-experts, and compiled to a loadable .h2p preset. Design principle: learn taste, compile contract.
一个「自然语言 → u-he Diva 合成器预设」生成系统。诸如 "warm analog bass with a soft attack" 的提示词先被转写为与合成器无关的语义意图（semantic intent），再由「按音色族条件化的专家混合模型（MoE）」映射为 881 维 Diva 参数向量，最终编译为可直接加载的 .h2p 预设。核心理念：学习品味，编译契约。
What I Built
two learned stages inside a deterministic shell
- • A semantic-IR transformer front-end that maps free text to a validated, synth-agnostic intent, with a deterministic constraint pass that guarantees internal coherence.
- • A family-conditioned residual mixture-of-experts (6 experts, activity-masked) that predicts the 249-field / 881-dim Diva parameter vector — the mechanism behind a measured typicality gain (family deviation 1.105 → 0.763).
- • An attribution-aware compiler with prompt-constraint repair contracts, plus a deterministic, admin-calibrated preference judge (78% → 86.4% agreement) extended with an offline audio term via a headless Diva render harness.
- • 语义 IR 转写前端：将自由文本映射为经校验、与合成器无关的意图，并以确定性约束层保证内部一致性。
- • 按音色族条件化的残差专家混合模型（6 专家，activity-mask）：预测 249 字段 / 881 维 Diva 参数向量——正是「音色典型度」提升（family deviation 1.105 → 0.763）背后的机制。
- • 归因感知编译器 + 提示约束修复契约；以及经管理员校准的确定性偏好判别器（一致率 78% → 86.4%），并通过无头 Diva 渲染管线扩展出离线音频项。
Tech Stack
core
Explore
product site & source