Third-party cyber evaluations involving OpenAI models
openai.com·14h ago
AI Research
arXiv CS.AI·Jun 10·5 min read
Interpreting and Steering a Text-to-Speech Language Model with Sparse Autoencoders
TL;DR
Text-to-speech (TTS) systems lack interpretability in how they generate speech from text. Sparse autoencoders (SAE) were developed to analyze and control features in the CosyVoice3 language model.
✦ Why It Matters
Engineers can use sparse autoencoders to gain interpretability and control over text-to-speech systems.
Key Takeaways
- Implement sparse autoencoders for feature analysis in TTS systems.
- Explore causal manipulation of speech features for better control.
- Utilize interpretable features to enhance user experience in TTS.
- Investigate the implications of feature control on speech synthesis quality.
Full Summary
Text-to-speech (TTS) systems rely on language models to convert text into spoken words, but understanding how these models represent speech features is limited. To address this, BatchTopK sparse autoencoders (SAE) were trained on the CosyVoice3 language model, creating a modality-aware pipeline that identifies the source of each feature—whether from text, speech, or both.
The resulting features were interpretable and included elements like phonemes, laughter, and speaker gender. By navigating the SAE latent space, researchers demonstrated that these features could be causally manipulated, achieving a laughter probability increase from 0.02 to 0.79 and successfully flipping the perceived gender of the speaker.
This work highlights the dual role of SAE features as both interpretable elements and control mechanisms for TTS synthesis, paving the way for more nuanced speech generation. The findings suggest that engineers can leverage these techniques to enhance TTS systems with greater control over speech attributes.