CosyVoice

Alibaba

Structures found in this family (2)

Linear Subspace, Linear Direction

By model (2)

CosyVoice2

Linear Subspace

CosyVoice3 (Qwen2.5-0.5B backbone) · 0.5B

Linear Direction

Observations (2)

CosyVoice2's LM module has a clean composable emotion subspace

SAE feature modality composition shifts by depth in a TTS LM and steers speech

Papers

A Geometric Perspective on Composable Emotion Steering in Text-to-Speech Models (2026), Interpreting and Steering a Text-to-Speech Language Model with Sparse Autoencoders (2026)