Segmentwise Pruning in Audio-Language Models
Abstract
While recent audio-language models excel at long-form tasks, their computational costs rise steeply with sequence length. Drawing on token pruning successes in vision-language models, we investigate their effectiveness in the audio domain. We further propose an improved, lightweight pruning strategy that explicitly leverages the time dimension for greater efficiency. While retaining only a quarter of the initial tokens, our approach results in a relative maximum decrease of 2% in CIDEr on Clotho v2 and a relative maximum decrease of 4% in accuracy on MMAU.
Fields of Study & Subjects
Music and Audio Processing
Speech Recognition and Synthesis
Speech and Audio Processing
Publication details
- Journal / source
- ICASSP 2026 - 2026 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)
- Publisher
- IEEE
- Pages
- 15322–15326
- Publication date
- 21 April 2026