Daily Arxiv

This is a page that curates AI-related papers published worldwide.
All content here is summarized using Google Gemini and operated on a non-profit basis.
Copyright for each paper belongs to the authors and their institutions; please make sure to credit the source when sharing.

DRASP: A Dual-Resolution Attentive Statistics Pooling Framework for Automatic MOS Prediction

Created by
  • Haebom

Author

Cheng-Yeh Yang, Kuan-Tang Huang, Chien-Chun Wang, Hung-Shin Lee, Hsin-Min Wang, Berlin Chen

Outline

This paper proposes a novel pooling mechanism, the Dual-Resolution Attentive Statistics Pooling (DRASP) framework, for MOS prediction, a speech quality assessment metric. To overcome the limitations of existing pooling methods, which tend to focus on global or frame-by-frame analysis and overlook complementary perceptual insights, DRASP integrates global statistical summaries with fine-grained analyses of key segments. This simultaneously captures both the overall structural context and important local details, yielding more accurate and robust representations. Extensive experiments on diverse datasets (MusicEval, AES-Natural), MOS prediction backbones (CLAP-based models, AudioBox-Aesthetics), and speech generation systems demonstrate the effectiveness and superior generalization performance of DRASP, improving the system-level Spearman correlation coefficient (SRCC) by 10.39% compared to the average pooling method.

Takeaways, Limitations

•
Takeaways:
◦
A novel pooling mechanism, DRASP, is proposed to effectively handle variable-length audio features.
◦
Improving MOS prediction performance by simultaneously considering global and local information.
◦
Demonstrated excellent performance and generalization across diverse datasets and models.
◦
Significant performance improvement compared to average pooling (SRCC improvement of 10.39%)
•
Limitations:
◦
Lack of analysis of DRASP's computational complexity and efficiency.
◦
Further validation of generalization performance across various audio quality degradation types is needed.
◦
Lack of detailed description of DRASP's parameter optimization strategy.
👍
Made with Slashpage