Research map / Large language models / Language, safety and interpretability
Sparse Autoencoders for Interpretable Emotion Control in Text-to-Speech
Hongfei Du, Jiacheng Shi, Sidi Lu, Gang Zhou, Ashley Gao
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| steer 6 selectivity-ranked SAE latents instead of 6 random SAE latents (ablation) | SAE-Emotion on IndexTTS2 · Emotion induction, happiness | Emo-SIM: 0.813 -> 0.885 | Improved |
| use sparse latent steering instead of dense mean-difference global steering (ablation) | IndexTTS2 · Emotion induction, anger/happiness/sadness | Emo-SIM: 0.910 -> 0.912 (anger); 0.879 -> 0.885 (happiness); 0.876 -> 0.880 (sadness) | Improved |
| use sparse latent steering instead of global steering for emotion suppression (ablation) | IndexTTS2 · Emotion suppression, anger | Emo-SIM: 0.915 -> 0.939 | Improved |
| use sparse latent steering instead of global steering (ablation) | IndexTTS2 · Emotion suppression, WER (anger) | WER: 2.6 -> 2.8 | Hurt |
| use sparse latent steering instead of global steering at a high steering scale (ablation) | IndexTTS2 · Strong steering robustness | Mean WER: 2.86% -> 0.57% | Improved |
| select latents by sentence-level selectivity rather than magnitude (ablation) | SAE-Emotion · Anger induction | Emo-SIM: 0.822 -> 0.912 | Improved |
| select latents by sentence-level selectivity rather than token-level criterion (ablation) | SAE-Emotion · Happiness induction | Emo-SIM: 0.811 -> 0.885 | Improved |
| increase SAE latent dimension from 4096 to 10,240 (ablation) | SAE on IndexTTS2 · SAE training | emotion interpretability | Hurt |
| increase number of steered latent features from 1 to 3 to 6 (ablation) | SAE-Emotion · Anger and sadness prototype alignment | emotion alignment | Improved |
| steer one emotion-related latent feature positively | IndexTTS2 · Paired acoustic analysis | mean F0 (Hz): 167.99 -> 191.10 | Improved |
| steer one emotion-related latent feature positively | IndexTTS2 · Paired acoustic analysis | duration (s): 1.685 -> 1.662 | No clear effect |
| increase steering scale alpha from -60 to +60 (ablation) | Happiness latent feature · Emotion prototype similarity | cosine similarity: 0.77 -> 0.86 | Improved |
Closest papers by meaning
- Word-Level Emotional Expression Control in Zero-Shot Text-to-Speech Synthesis (NeurIPS 2025 · similarity 0.88)
- EmotionThinker: Prosody-Aware Reinforcement Learning for Explainable Speech Emotion Reasoning (ICLR 2026 · similarity 0.87)
- EmoReg: Directional Latent Vector Modeling for Emotional Intensity Regularization in Diffusion-based Voice Conversion (AAAI 2025 · similarity 0.86)
- VowelPrompt: Hearing Speech Emotions from Text via Vowel-level Prosodic Augmentation (ICLR 2026 · similarity 0.86)
- SAEBench: A Comprehensive Benchmark for Sparse Autoencoders in Language Model Interpretability (ICML 2025 · similarity 0.86)
- Improving Sparse Decomposition of Language Model Activations with Gated Sparse Autoencoders (NeurIPS 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
