Research map / Image generation / Autoregressive and latent generation
ParaMETA: Towards Learning Disentangled Paralinguistic Speaking Styles Representations from Speech
Haowei Lou, Hye-young Paik, Wen Hu, Lina Yao
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add META-space regularization (ablation) | ParaMETA (Transformer) · Speaking-style classification: Emotion | balanced accuracy (%): 44.2 -> 50.1 | Improved |
| add META-space regularization (ablation) | ParaMETA (Transformer) · Speaking-style classification: Age | balanced accuracy (%): 26.1 -> 29.7 | Improved |
| add META-space regularization (ablation) | ParaMETA (LSTM) · Speaking-style classification: Emotion | balanced accuracy (%): 51.6 -> 49.0 | Hurt |
| add META-space regularization (ablation) | ParaMETA (LSTM) · Speaking-style classification: Age | balanced accuracy (%): 36.3 -> 32.8 | Hurt |
| add META-space regularization (ablation) | ParaMETA (QFormer) · Speaking-style classification: Age | balanced accuracy (%): 38.2 -> 35.8 | Hurt |
| add META-space regularization (ablation) | ParaMETA (CNN) · Speaking-style classification: Gender | balanced accuracy (%): 61.8 -> 64.2 | Improved |
| add META-space regularization across backbones and tasks (ablation) | ParaMETA · Speaking-style classification (16 backbone-task settings) | balanced accuracy (%): improves performance in 8 of 16 backbone-task settings | Mixed |
| replace cross-entropy multi-task training with ParaMETA (w r) | Transformer encoder · Speaking-style classification: Emotion | balanced accuracy (%): 35.0 -> 50.1 | Improved |
| replace cross-entropy multi-task training with ParaMETA (w r) | LSTM encoder · Speaking-style classification: Age | balanced accuracy (%): 23.5 -> 32.8 | Improved |
| replace cross-entropy multi-task training with ParaMETA (w r) | LSTM encoder · Speaking-style classification: Emotion | balanced accuracy (%): 61.3 -> 49.0 | Hurt |
| replace CLAP-style text-speech contrastive objective with ParaMETA (w r) | Transformer encoder · Speaking-style classification: Gender | balanced accuracy (%): 39.4 -> 78.4 | Improved |
| replace CLAP-style objective with ParaMETA (w r) | QFormer encoder · Speaking-style classification: Language | balanced accuracy (%): 48.6 -> 78.9 | Improved |
Closest papers by meaning
- A Variational Framework for Improving Naturalness in Generative Spoken Language Models (ICML 2025 · similarity 0.85)
- Leviathan: Decoupling Input and Output Representations in Language Models (NeurIPS 2026 · similarity 0.84)
- Emotional Face-to-Speech (ICML 2025 · similarity 0.84)
- Dynamic-SUPERB Phase-2: A Collaboratively Expanding Benchmark for Measuring the Capabilities of Spoken Language Models with 180 Tasks (ICLR 2025 · similarity 0.84)
- Paralinguistics-Aware Speech-Empowered Large Language Models for Natural Conversation (NeurIPS 2024 · similarity 0.84)
- SpeechAlign: Aligning Speech Generation to Human Preferences (NeurIPS 2024 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
