Interpreting and Steering a Text-to-Speech Language Model with Sparse Autoencoders
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.
The following analysis is generated by AI and may contain errors. It is separate from the authors’ abstract and is not peer review.
Research summary
The experimental evaluation is comprehensive and well-structured, providing both quantitative and qualitative insights. 1. **Layer Sweep Analysis**: Figure 1 effectively illustrates the shift in feature modality composition and explained variance across layers. This provides valuable insights into the LM's internal processing, showing a transition from mixed/text-heavy to audio-heavy features in middle layers, and a surprising reversion to text-modal in the final hidden state. This analysis is novel for TTS LMs. 2. **Auto-Interp Quality**: Figure 2 provides quantitative evaluation of the auto-interpretation quality, showing text-modal labels are most verifiable (AUROC 0.921), followed by audio-modal (0.653), and mixed (0.558). While the mixed feature scores are lower, they are still above chance, and the qualitative examples (Table 1) show that even some mixed features are clean. 3. **Qualitative Examples**: Table 1 presents compelling examples of interpretable features, spanning phonemes, laughter, accent prompts, speaker gender, words, and even sub-lexical patterns. These examples strongly support the claim of interpretability. 4. **Feature Steering**: This is the highlight of the results. The paper demonstrates *causal control* over three distinct speech properties: * Laughter probability (0.015 to 0.791). * Perceived speaker gender (wav2vec2 P(male) from 0.629 to 0.944 or 0.063). * Speech rate (voiced duration from 3.96s to 10.57s or 2.75s). These results are highly impactful, showing that the identified features are not just descriptive but functional. The separation of gender steering by prompt speaker's original gender (Figure 4) adds further robustness. 5. **Concept Probing Experiments**: The appendix details supervised probes for laughter, emotion, and accent, showing that these concepts are linearly decodable early in the network and that SAE-latent probes closely track raw
Methodology
The methodology is a well-executed adaptation of established mechanistic interpretability techniques (Sparse Autoencoders, LLM-based auto-interpretation) to a novel and challenging domain: the residual stream of a generative Text-to-Speech (TTS) Language Model. The core innovation lies in the "modality-aware auto-interp pipeline." This involves: 1. **SAE Training**: Training BatchTopK SAEs on the Qwen2.5-0.5B backbone of CosyVoice3, a standard and robust approach for decomposing activations. The choice of dictionary size and sparsity (d=16384, k=50) is reasonable for a model of this size. 2. **Evidence Extraction**: A crucial step that correctly identifies whether an activation occurs in the text-prefix or speech-token segment. This allows for modality-specific evidence (marked text window vs. 1-second audio clip). 3. **Feature Modality Tagging**: A simple yet effective heuristic (speech fraction > 0.8 for audio-modal, < 0.2 for text-modal, otherwise mixed) to categorize features. This is essential for guiding the auto-interpretation process. 4. **Automatic Labeling**: Using Gemini 3.0 Pro with modality-aware prompts is a clever way to leverage powerful LLMs for interpretation. The prompts are carefully designed to elicit specific descriptions based on the evidence type. 5. **Detection-Style Evaluation**: Adapting the protocol from text-only to mixed text/audio evidence, using rank-held-out examples and AUROC/balanced accuracy, provides a quantitative measure of label quality. 6. **SAE Feature Steering**: This is the most impactful methodological contribution. Instead of directly adding residual vectors, the intervention occurs in the SAE latent space, modifying selected feature activations and then decoding back to the residual stream. This demonstrates the *causal* role of the identified features, moving beyond mere correlation. The intervention mechanism is clearly described, ensuring locality to the SAE feature subspace.
Results & evaluation
The experimental evaluation is comprehensive and well-structured, providing both quantitative and qualitative insights. 1. **Layer Sweep Analysis**: Figure 1 effectively illustrates the shift in feature modality composition and explained variance across layers. This provides valuable insights into the LM's internal processing, showing a transition from mixed/text-heavy to audio-heavy features in middle layers, and a surprising reversion to text-modal in the final hidden state. This analysis is novel for TTS LMs. 2. **Auto-Interp Quality**: Figure 2 provides quantitative evaluation of the auto-interpretation quality, showing text-modal labels are most verifiable (AUROC 0.921), followed by audio-modal (0.653), and mixed (0.558). While the mixed feature scores are lower, they are still above chance, and the qualitative examples (Table 1) show that even some mixed features are clean. 3. **Qualitative Examples**: Table 1 presents compelling examples of interpretable features, spanning phonemes, laughter, accent prompts, speaker gender, words, and even sub-lexical patterns. These examples strongly support the claim of interpretability. 4. **Feature Steering**: This is the highlight of the results. The paper demonstrates *causal control* over three distinct speech properties: * Laughter probability (0.015 to 0.791). * Perceived speaker gender (wav2vec2 P(male) from 0.629 to 0.944 or 0.063). * Speech rate (voiced duration from 3.96s to 10.57s or 2.75s). These results are highly impactful, showing that the identified features are not just descriptive but functional. The separation of gender steering by prompt speaker's original gender (Figure 4) adds further robustness. 5. **Concept Probing Experiments**: The appendix details supervised probes for laughter, emotion, and accent, showing that these concepts are linearly decodable early in the network and that SAE-latent probes closely track raw
Author of this paper? Send a correction.