> When text and speech tokens dance in the same residual stream, can you tell who is leading? A new paper inserts sparse autoencoders (SAEs) into the backbone of a TTS language model for the first time, revealing a layered trajectory from text prefix to audio commitment and back to text.
Why dissect a TTS model?
Modern text-to-speech systems increasingly rely on language models as backbones. CosyVoice3 uses Qwen2.5-0.5B to autoregressively generate speech from mixed sequences of text prefixes and discrete speech tokens. But since text and audio share the same residual stream, no one had systematically answered: what does each layer encode—text syntax/semantics, acoustic/prosodic properties, or cross-modal mappings? While SAE interpretability research abounds for text LLMs, mechanistic interpretability of generative TTS models has been nearly blank. This paper (arXiv:2606.10029, Nikita Koriagin et al., T-Tech AI Foundation) is the first full-layer SAE scan of CosyVoice3's backbone, with automated labeling and causal steering experiments.
Experimental setup
| Property | Spec |
|:---|:---|
| LM backbone | Qwen2.5-0.5B, hidden dim 896 |
| Layers | 28 (analyzing layers 0–23) |
| Speech tokens | 25 Hz, autoregressive |
| Sequence layout | [sos \| instruct \| text \| task \| speech] |
| Pipeline | Text/prompt → Qwen2.5 LLM → discrete speech tokens → DiT flow matching → HiFi-GAN vocoder |
SAE configuration: BatchTopK variant, dictionary size 16,384, 50 active features per token, trained on ~250M tokens from the Emilia dataset, with reconstruction + sparsity + auxiliary dead-feature losses.
A surprising layered trajectory
- Early/middle layers (0–14): mixed + audio dominant; text modality only 12–33%; mixed features peak at 47.3% at layer 12.
- Late layers (16–20): an "audio commitment zone." Audio features hold at 74–76% while mixed features collapse from 40.9% to 4.1%. The model "decides" what the speech will sound like here. Reconstruction is hardest at layer 20 (EV 0.82, the lowest), then rebounds to 0.945 at layer 23.
- Final layer (23): a "text rebound." Text modality sharply returns to 83.1%—the residual stream re-projects onto a text-vocabulary-aligned subspace before the output head.
- Text features: AUROC 0.921 (e.g., feature 1376 = "British", feature 1443 = substring "ang", feature 1330 = four-digit years)
- Audio features: AUROC 0.653 (e.g., human laughter, screams and heavy breathing, the /k/ plosive)
- Mixed features: AUROC 0.558 (e.g., stuttering/disfluency markers, "middle" appearing in both text and speech)
- Laughter (feature 14834): mean laughter probability rises from 0.015 to 0.791 at α = +60 (53x).
- Gender (feature 11402): P(male) goes from 0.629 baseline to 0.944 (α = -50) or 0.063 (α = +50), flipping perceived speaker gender regardless of prompt.
- Speech rate (feature 3024): voiced duration scales from 3.96 s to 10.57 s (2.67x slower) or 2.75 s (69%, faster)—while preserving spoken content.
- Koriagin et al., "Interpreting and Steering a Text-to-Speech Language Model with Sparse Autoencoders", arXiv:2606.10029 (2026)
- Model: CosyVoice3 (Qwen2.5-0.5B), 25 Hz discrete speech tokens
- Data: Emilia (~250M tokens), VocalSound, ESD, VCTK, LJSpeech
- Labeling: Gemini 3.0 Pro; probing: binary logistic regression with 5-fold cross-validation
Modality-aware automated labeling
Features were classified as text (activating mainly on text prefixes), audio (mainly on speech tokens), or mixed, then labeled by Gemini 3.0 Pro using text context, 1-second audio snippets, or both. Detection-style evaluation showed:
Probing: concepts exposed by layer 8
Linear probes show laughter, emotion, and accent reach ROC-AUC > 0.99 by layer 8, with SAE latents closely tracking raw residual probing. Top-1 single-feature tests yield ROC-AUC ≈ 0.93 (laughter 0.929, emotion 0.928, accent 0.925)—concepts are localized in a few interpretable dictionary directions rather than fully distributed.
Causal steering: from understanding to control
Interventions encode residual vectors at chosen speech-token positions, perturb specific feature activations, and decode back—without touching the input text:
Limitations
The authors candidly list six: single model (CosyVoice3-0.5B); circular evaluation (Gemini as both labeler and grader); partial layer coverage for explanation scores; inability to localize sub-token (40 ms) acoustic onsets at 25 Hz; negative sampling that tests specificity but not neighborhood confusion; and unexplored boundaries of the steering coefficient α.
Conclusion
The paper's key message: SAE features are both interpretability objects and practical control directions. Unlike prompt-based control ("say it cheerfully"), SAE steering operates directly on the residual stream—precise, quantifiable, and pluggable. Future work includes scaling to larger TTS models, human evaluation of labels, exploring more controllable dimensions (emotion intensity, pitch), and real-time integration into production TTS systems.
References