Problem
Research questions and friction points this paper is trying to address.
paralinguistic variation
speech language model
latent geometry
content decomposition
latent speech control
ParaGeo: Decomposing Paralinguistic Variation into a Shared Latent Geometry
🤖 AI Summary
This study addresses the challenge of disentangling paralinguistic attributes from textual content in speech, which are inherently coupled. Within a frozen large speech model, this work fixes listen-and-read prompts to match content, thereby isolating paralinguistic variation and constructing a shared low-dimensional geometric space to quantify its structure. A method based on key/value representation projection is proposed, integrating centering, dimensionality reduction, and global calibration basis fitting to achieve effective separation and controllable intervention of content and paralinguistic properties. Experimental results demonstrate that the proposed approach significantly outperforms baselines in centroid accuracy and substantially improves cross-content cosine similarity, validating both the effectiveness of the shared coordinate representation and the reproducibility of the control directions.
📝 Abstract
Speech delivery varies with both the requested paralinguistic attribute and the linguistic content. We introduce ParaGeo, a matched-content decomposition of paralinguistic variation in a frozen speech language model. Synthesized audio tokens are replayed with a fixed listening prompt; pooled key/value (K/V) representations are centered and projected into a shared low-dimensional space. Our GLM-4-Voice probe spans 80 requested controls from 12 benchmark families across eight sentences. With a globally fitted calibration basis, content-held-out centroid accuracy using this basis is 9.49% versus a 1.25% permutation baseline; same-label cross-content cosine similarity is 0.285 versus 0.017, and both conditional permutation tests yield p = 0.001. A separate ten-scenario, six-style probe reveals reproducible contrast directions across scenarios. Static, additive, and temporal interventions produce attribute-, layer-, and schedule-dependent response profiles. These results provide a shared coordinate representation for measuring paralinguistic structure and an empirical starting point for latent speech control. Code is available at https://github.com/yuhanlydia/ParaGeo.
Innovation
Methods, ideas, or system contributions that make the work stand out.
Paralinguistic decomposition
Shared latent geometry
Frozen speech language model
K/V representations
Latent speech control
🔎 Similar Papers
No similar papers found.
