mlsanityData to 14 Oct 2026

A Geometric Perspective on Composable Emotion Steering in Text-to-Speech Models

arXiv 2607.00946 ↗First posted 1 Jul 2026cs.SD cs.LGcited by 3

On the map: Mechanistic interpretability → Activation steering methods

Abstract (arXiv)

While prior work has explored emotion control in hybrid text-to-speech systems, the geometric properties of these modules, and their implications for steerability, remain poorly understood. We present the first comparative study of speech language model (SLM) and conditional flow-matching (CFM) modules as activation steering sites for mixed emotion speech synthesis. We first characterize emotion representations using linear probing and local intrinsic dimensionality (LID), and then evaluate single-site and joint steering for mixed-emotion synthesis. Our results show that SLM offers a clean, low-dimensional emotion-specific subspace with strong speaker--emotion disentanglement, while CFM exhibitspoor cross-speaker generalization due to speaker--emotion entanglement. Joint steering increases emotion intensity but degrades proportional control and speech quality on in-distribution data. These findings provide practical guidance for multi-site activation steering in hybrid TTS systems and highlight the importance of representation geometry in controllable speech generation.

Key facts. Cited by 3 papers in the map (data through 14 Oct 2026). Placed in Activation steering methods, one of 13 directions in Mechanistic interpretability. Builds on 23 of its 29 references that are in the map. First posted on arXiv 1 Jul 2026.

Builds on

23 of 29 references.

Closest ancestors · 4

Same direction, newest first

  1. 2026CoCoEmo: Composable and Controllable Human-Like Emotional TTS via Activation Steering3 Feb
  2. 2025EmoSteer-TTS: Fine-Grained and Training-Free Emotion-Controllable Text-to-Speech via Activation Steering5 Aug
  3. 2023Steering Llama 2 via Contrastive Activation Addition9 Dec
  4. 2023Steering Language Models With Activation Engineering20 Aug

Other references · 19

  1. 2022Flow Matching for Generative Modelingcited by 5,702
  2. 2022Robust Speech Recognition via Large-Scale Weak Supervisioncited by 4,057
  3. 2021WavLM: Large-Scale Self-Supervised Pre-Training for Full Stack Speech Processingcited by 1,657
  4. 2016Understanding intermediate layers using linear classifier probescited by 1,353
  5. 2023Representation Engineering: A Top-Down Approach to AI Transparencycited by 1,336
  6. 2020Null It Out: Guarding Protected Attributes by Iterative Nullspace Projectioncited by 455
  7. 2024Seed-TTS: A Family of High-Quality Versatile Speech Generation Modelscited by 300
  8. 2024CosyVoice 2: Scalable Streaming Speech Synthesis with Large Language Modelscited by 288
  9. 2023emotion2vec: Self-Supervised Pre-Training for Speech Emotion Representationcited by 205
  10. 2020Seen and Unseen emotional style transfer for voice conversion with a new emotional speech datasetcited by 121
  11. 2023The geometry of hidden representations of large transformer modelscited by 112
  12. 2022PromptTTS: Controllable Text-to-Speech with Text Descriptionscited by 110
7 more
  1. 2025IndexTTS2: A Breakthrough in Emotionally Expressive and Duration-Controlled Auto-Regressive Zero-Shot Text-to-Speechcited by 90
  2. 2024Controlling Language and Diffusion Models by Transporting Activationscited by 68
  3. 2025EmoVoice: LLM-based Emotional Text-To-Speech Model with Freestyle Text Promptingcited by 34
  4. 2024Emo-DPO: Controllable Emotional Speech Synthesis through Direct Preference Optimizationcited by 32
  5. 2022Speech Synthesis with Mixed Emotionscited by 28
  6. 2025Representation Engineering for Large-Language Models: Survey and Research Challengescited by 21
  7. 2024EmoSphere++: Emotion-Controllable Zero-Shot Text-to-Speech via Emotion-Adaptive Spherical Vectorcited by 19

Built on by

Cited by year (2026 to 14 Oct)

Citing papers per year: 2026: 3.

3 papers.

  1. 2026EmoRES-TTS: Residual-Enhanced Vector Steering for Emotional Speech Generationcited by 0
  2. 2026ParaGeo: Decomposing Paralinguistic Variation into a Shared Latent Geometrycited by 0
  3. 2026Steering Follows Geometry, Not Labels: Emotion Directions in a Full-Duplex Speech Modelcited by 0

Similar, not cited

Close in wording, no citation either way

  1. 2026DUET: Unified Dual-Space Emotion Control for Diffusion and Flow-Matching Driven Text-to-Speech
  2. 2026Steerspeech: Activation Steering For Emotion Control In Generated Speech
  3. 2024Hierarchical Control of Emotion Rendering in Speech Synthesis
  4. 2026EmoShift: Lightweight Activation Steering for Enhanced Emotion-Aware Speech Synthesis
  5. 2026Tracing a Sparse Emotion-Control Circuit in LLM-Based Text-to-Speech
  6. 2025TTS-CtrlNet: Time varying emotion aligned text-to-speech generation with ControlNet
  7. 2026Sparse Autoencoders for Interpretable Emotion Control in Text-to-Speech
  8. 2023AffectEcho: Speaker Independent and Language-Agnostic Emotion and Affect Transfer for Speech Synthesis