pith. sign in

arxiv: 2606.10029 · v1 · pith:3VDZDIHWnew · submitted 2026-06-08 · 💻 cs.LG · cs.AI· cs.CL

Interpreting and Steering a Text-to-Speech Language Model with Sparse Autoencoders

Nikita Koriagin , Georgii Aparin , Nikita Balagansky , Daniil Gavrilov This is my paper

Pith reviewed 2026-06-27 17:08 UTC · model grok-4.3

classification 💻 cs.LG cs.AIcs.CL
keywords sparse autoencoderstext-to-speechlanguage modelsinterpretabilitysteeringcausal featuresmultimodal modelsresidual stream
0
0 comments X

The pith

Sparse autoencoders on a TTS language model recover features that can be steered to control laughter, speaker gender, and speech rate.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper trains BatchTopK sparse autoencoders on the language model backbone of a text-to-speech system in which text and speech tokens share one residual stream. It adds a modality-aware auto-interpretation pipeline that labels each feature according to whether it fires on text prefixes, short speech clips, or both. The resulting features correspond to concrete speech attributes such as phonemes, laughter, accents, and speaker gender. Interventions that add or subtract these features produce measurable causal effects, including raising laughter probability from 0.02 to 0.79, flipping perceived gender, and adjusting speech rate without changing the spoken words. The work therefore treats the same SAE latents as both interpretability objects and practical control directions for synthesis.

Core claim

We train BatchTopK sparse autoencoders on the LM backbone of CosyVoice3 and introduce a modality-aware auto-interp pipeline that labels each feature from where it fires—text-prefix context, 1-second speech clips, or both. The recovered features are interpretable, spanning phonemes, laughter, accent prompts and speaker gender. Steering through the SAE latent space shows these features are causal rather than merely descriptive: targeted interventions raise laughter probability from 0.02 to 0.79, flip perceived speaker gender, and control speech rate while preserving spoken content. SAE features thus serve both as interpretability objects and as control directions for TTS synthesis.

What carries the argument

BatchTopK sparse autoencoders paired with a modality-aware auto-interpretation pipeline that tags each latent by its activation context in text or speech.

If this is right

  • Targeted addition of an SAE feature can raise laughter probability from 0.02 to 0.79 while leaving spoken content unchanged.
  • Subtraction or addition of gender-related features can flip perceived speaker gender in the output audio.
  • Adjustment of rate-related features can alter speech speed without altering the words produced.
  • The same latents function simultaneously as readable descriptions and as editable control vectors for synthesis.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same pipeline could be applied to other TTS language models that interleave text and speech tokens to test whether comparable causal features emerge.
  • Multiple SAE features might be combined in a single forward pass to achieve simultaneous control over several attributes such as emotion and accent.
  • If the recovered features remain stable across different speakers and prompts, they could serve as a lightweight interface for editing generated speech after the initial decoding step.

Load-bearing premise

The assumption that the SAE features recovered by the modality-aware auto-interp pipeline are the correct level of abstraction and that the observed steering effects are not accompanied by unmeasured degradations in overall speech quality or coherence.

What would settle it

A controlled listening test in which laughter-directed interventions are applied yet the measured laughter probability remains near baseline or listener ratings of coherence drop measurably.

Figures

Figures reproduced from arXiv: 2606.10029 by Daniil Gavrilov, Georgii Aparin, Nikita Balagansky, Nikita Koriagin.

Figure 1
Figure 1. Figure 1: Layer sweep across the Qwen2.5-0.5B backbone. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Held-out auto-interp scores for layer-20 fea [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Gender steering split by prompt-speaker gender. Feature 11402 shifts both male- and female￾prompted generations toward the target gender. 4.4 Feature Steering We finally test whether interpreted SAE features can be used as causal controls for synthesis. We steer three layer-20 features with labels, shown in [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Steering effects for laughter, speaker-gender [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: CosyVoice-aware view of our modality-aware SAE interpretation pipeline. CosyVoice3 synthesizes speech [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Rank-held-out auto-interp scores across completed layers. Text-modal labels remain consistently easiest [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Raw-residual probe ROC-AUC as a function of layer for laughter, emotion, and accent. Emotion and [PITH_FULL_IMAGE:figures/full_fig_p010_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Top-1 SAE feature ROC-AUC as a function of layer. This monosemanticity test asks whether a single [PITH_FULL_IMAGE:figures/full_fig_p010_8.png] view at source ↗
read the original abstract

Language models increasingly serve as the backbone of text-to-speech (TTS) systems, yet we understand little about the representations they build when text and generated speech tokens share a single residual stream. We train BatchTopK sparse autoencoders on the LM backbone of CosyVoice3 and introduce a modality-aware auto-interp pipeline that labels each feature from where it fires-text-prefix context, 1-second speech clips, or both. The recovered features are interpretable, spanning phonemes, laughter, accent prompts and speaker gender. Steering through the SAE latent space shows these features are causal rather than merely descriptive: targeted interventions raise laughter probability from 0.02 to 0.79, flip perceived speaker gender, and control speech rate while preserving spoken content. SAE features thus serve both as interpretability objects and as control directions for TTS synthesis.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 0 minor

Summary. The manuscript trains BatchTopK sparse autoencoders on the residual stream of the LM backbone in CosyVoice3. It introduces a modality-aware auto-interp pipeline that labels features according to activation context (text-prefix, 1-second speech clips, or both). The recovered features are interpretable, covering phonemes, laughter, accents, and speaker gender. Steering interventions in the SAE latent space are reported to demonstrate causality, raising laughter probability from 0.02 to 0.79, flipping perceived speaker gender, and controlling speech rate while preserving spoken content. The authors conclude that SAE features function as both interpretability objects and control directions for TTS synthesis.

Significance. If the steering results hold without unmeasured quality trade-offs, the work extends SAE-based interpretability to multimodal TTS LMs and supplies concrete quantitative intervention outcomes (laughter probability shift of 0.02 to 0.79) that could support more precise synthesis control. The modality-aware labeling approach is a methodological contribution for mixed-modality activations.

major comments (2)
  1. [Abstract] Abstract: The central claim that SAE steering demonstrates causality (rather than mere description) and 'preserves spoken content' is not accompanied by any quantitative post-steering metrics on naturalness, intelligibility, prosody consistency, or coherence (e.g., no MOS, WER, or artifact rates). This is load-bearing for the causality interpretation, because large attribute shifts could be side-effects of general disruption.
  2. [Results] Results (steering experiments): No baselines or statistical controls are described for the reported effect sizes relative to random latent interventions or non-SAE steering vectors. Without these, it is unclear whether the observed changes (laughter 0.02→0.79, gender flip, rate control) are specifically attributable to the recovered SAE features.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their careful reading and constructive feedback. The comments correctly identify gaps in the quantitative support for our steering claims. We address each point below and commit to revisions that will strengthen the manuscript.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The central claim that SAE steering demonstrates causality (rather than mere description) and 'preserves spoken content' is not accompanied by any quantitative post-steering metrics on naturalness, intelligibility, prosody consistency, or coherence (e.g., no MOS, WER, or artifact rates). This is load-bearing for the causality interpretation, because large attribute shifts could be side-effects of general disruption.

    Authors: We agree that the absence of post-intervention metrics on naturalness and content preservation weakens the causality interpretation. In the revised manuscript we will add WER computed via an off-the-shelf ASR model on steered versus baseline utterances, together with human MOS ratings for naturalness and artifact presence. These numbers will be reported for all steering experiments. revision: yes

  2. Referee: [Results] Results (steering experiments): No baselines or statistical controls are described for the reported effect sizes relative to random latent interventions or non-SAE steering vectors. Without these, it is unclear whether the observed changes (laughter 0.02→0.79, gender flip, rate control) are specifically attributable to the recovered SAE features.

    Authors: The referee is correct that specificity to the learned SAE directions has not been demonstrated. We will add, in the revised results section, (i) steering with random vectors of matched norm, (ii) direct activation patching without the SAE, and (iii) bootstrap or permutation tests comparing the observed effect sizes against these controls. The new tables will quantify how much of the reported attribute change is attributable to the SAE features themselves. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical intervention study with measured effects

full rationale

The paper trains BatchTopK SAEs on the CosyVoice3 LM backbone, applies a modality-aware auto-interp pipeline to label features, and then performs targeted steering interventions whose effects (laughter probability shift, gender flip, rate control) are directly measured on held-out outputs. No equations, predictions, or uniqueness claims reduce any reported quantity to the SAE training loss or fitted parameters by construction. The central causal claim rests on external empirical measurements rather than self-referential definitions or self-citation chains.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Only the abstract is available; no explicit free parameters, axioms, or invented entities are stated.

pith-pipeline@v0.9.1-grok · 5686 in / 1019 out tokens · 15387 ms · 2026-06-27T17:08:00.562220+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

16 extracted references · 7 canonical work pages · 4 internal anchors

  1. [1]

    Sparse Autoencoders Find Highly Interpretable Features in Language Models

    Sparse autoencoders find highly interpretable features in language models , author=. arXiv preprint arXiv:2309.08600 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  2. [2]

    Anthropic, accessed , year=

    Scaling and evaluating sparse autoencoders , author=. Anthropic, accessed , year=

  3. [3]

    OpenAI blog , year=

    Language models can explain neurons in language models , author=. OpenAI blog , year=

  4. [4]

    Automatically interpreting millions of features in large language models.arXiv preprint arXiv:2410.13928,

    Automatically Interpreting Millions of Features in Large Language Models , author=. arXiv preprint arXiv:2410.13928 , year=

    work page arXiv

  5. [5]

    Gemma Scope: Open Sparse Autoencoders Everywhere All At Once on Gemma 2

    Gemma Scope: Open Sparse Autoencoders Everywhere All At Once on Gemma 2 , author=. arXiv preprint arXiv:2408.05147 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  6. [6]

    Scaling and evaluating sparse autoencoders

    Scaling and evaluating sparse autoencoders , author=. arXiv preprint arXiv:2406.04093 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  7. [7]

    arXiv preprint arXiv:2407.05361 , year=

    Emilia: An Extensive, Multilingual, and Diverse Speech Dataset for Large-Scale Speech Generation , author=. arXiv preprint arXiv:2407.05361 , year=

    work page arXiv

  8. [8]

    Du, Zhihao and Chen, Qian and Zhang, Shiliang and Hu, Kai and Lu, Heng and Yang, Yexin and Hu, Hangrui and Zheng, Siqi and Gu, Yue and Ma, Ziyang and others , journal=

  9. [9]

    ASRU , year=

    Layer-wise analysis of a self-supervised speech representation model , author=. ASRU , year=

  10. [10]

    AudioSAE: Towards Understanding of Audio-Processing Models with Sparse AutoEncoders , url=

    Aparin, Georgii and Sadekova, Tasnima and Rukhovich, Alexey and Yermekova, Assel and Kushnareva, Laida and Popov, Vadim and Kuznetsov, Kristian and Piontkovskaya, Irina , year=. AudioSAE: Towards Understanding of Audio-Processing Models with Sparse AutoEncoders , url=. doi:10.18653/v1/2026.eacl-long.149 , booktitle=

    work page doi:10.18653/v1/2026.eacl-long.149 2026

  11. [11]

    2024 , institution=

    Gemini: A Family of Highly Capable Multimodal Models , author=. 2024 , institution=

    2024

  12. [12]

    and Ellis, Daniel P

    Gemmeke, Jort F. and Ellis, Daniel P. W. and Freedman, Dylan and Jansen, Aren and Lawrence, Wade and Moore, R. Channing and Plakal, Manoj and Ritter, Marvin , booktitle=

  13. [13]

    Gong, Yuan and Chung, Yu-An and Glass, James , booktitle=

  14. [14]

    ICML , year=

    Robust speech recognition via large-scale weak supervision , author=. ICML , year=

  15. [15]

    Whisper Hallucination Detection and Mitigation via Hidden Representation Steering and Sparse AutoEncoders

    Whisper Hallucination Detection and Mitigation via Hidden Representation Steering and Sparse AutoEncoders , author=. arXiv preprint arXiv:2606.07473 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  16. [16]

    Xie, Tianxin and Yang, Shan and Li, Chenxing and Yu, Dong and Liu, Li , journal=