arXivDaily arXiv每日学术速递 周一至周五更新
arXiv周末暂无论文更新,休息一下吧,周末愉快~~
arXiv 2606.10029cs.LGcs.AIcs.CL

用稀疏自编码器解释和引导文本转语音语言模型

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

T-Tech;AI Foundation and Algorithm Lab(人工智能基础与算法实验室)

Nikita Koriagin, Georgii Aparin, Nikita Balagansky, Daniil Gavrilov

AI总结:

本文在CosyVoice3语言模型骨干上训练BatchTopK稀疏自编码器,发现特征可解释且因果可控,能操纵笑声、性别和语速。

摘要:

语言模型日益成为文本转语音(TTS)系统的骨干,但我们对其在文本和生成语音令牌共享单一残差流时构建的表示知之甚少。我们在CosyVoice3的语言模型骨干上训练BatchTopK稀疏自编码器,并引入一种模态感知的自动解释流水线,根据特征触发的位置——文本前缀上下文、1秒语音片段或两者——为每个特征打标签。恢复的特征是可解释的,涵盖音素、笑声、口音提示和说话者性别。通过SAE潜在空间进行引导表明,这些特征是因果性的而非仅仅是描述性的:有针对性的干预将笑声概率从0.02提高到0.79,翻转感知到的说话者性别,并在保持口语内容的同时控制语速。因此,SAE特征既可作为解释性对象,也可作为TTS合成的控制方向。

英文摘要:

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.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑