On Sequence Segmentation with overlapped Chunks in Machine Learning

Joel Rixen, Matthias Renz

Keywords: sequence segmentation, speech separation, source separation, audio super resolution, stft, signal processing

Abstract:

Operating on very long sequences can be problematic for many sequence modelling methods like Transformers or recurrent neural networks. To avoid this issue, long sequences are often split into smaller chunks instead.
For various reasons, these chunks typically are overlapped with each other which causes an increase in tensor size by however much the chunks are overlapping.

This paper attempts to find a better understanding on overlapped sequence chunks and what they accomplish. Specifically, the focus of this paper is on audio inputs in both the time and frequency domain. Previous models for speech separation and audio super resolution which use overlapped chunks are modified to allow for reduced or even removed overlaps which causes significant decreases in computational cost while maintaining accuracy.

Primary Area: applications to computer vision, audio, language, and other modalities

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 3691