Windowed SummaryMixing: An Efficient Fine-Tuning of Self-Supervised Learning Models for Low-resource Speech Recognition

Fuente: arXiv
Gespeichert in:
Bibliographische Detailangaben
Hauptverfasser: Menon, Aditya Srinivas, Tripathi, Kumud, Gohil, Raj, Wasnik, Pankaj
Format: Preprint
Veröffentlicht: 2026
Schlagworte:
Online-Zugang:
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
_version_ 1866915787038523392
author Menon, Aditya Srinivas
Tripathi, Kumud
Gohil, Raj
Wasnik, Pankaj
author_facet Menon, Aditya Srinivas
Tripathi, Kumud
Gohil, Raj
Wasnik, Pankaj
contents Self-supervised learning (SSL) has advanced speech processing but suffers from quadratic complexity due to self-attention. To address this, SummaryMixing (SM) has been proposed as a linear-time alternative that summarizes entire utterances using mean pooling but lacks sufficient local context. In this work, we introduce Windowed SummaryMixing (WSM), which enhances SM by integrating local neighborhood summaries alongside the global summary, maintaining efficiency while improving temporal dependencies. Additionally, we introduce a selective fine-tuning approach, replacing self-attention layers in SSL models with WSM blocks and fine-tuning only these blocks in low-resource settings. Our approach improves ASR performance while reducing peak VRAM usage by 40\% in the SSL models. WSM blocks have linear-time complexity with enhanced context awareness. Selectively replacing some attention layers reduces compute, memory, and latency, making it ideal for low-resource speech recognition.
format Preprint
id arxiv_https___arxiv_org_abs_2602_09043
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Windowed SummaryMixing: An Efficient Fine-Tuning of Self-Supervised Learning Models for Low-resource Speech Recognition
Menon, Aditya Srinivas
Tripathi, Kumud
Gohil, Raj
Wasnik, Pankaj
Audio and Speech Processing
Machine Learning
Sound
Self-supervised learning (SSL) has advanced speech processing but suffers from quadratic complexity due to self-attention. To address this, SummaryMixing (SM) has been proposed as a linear-time alternative that summarizes entire utterances using mean pooling but lacks sufficient local context. In this work, we introduce Windowed SummaryMixing (WSM), which enhances SM by integrating local neighborhood summaries alongside the global summary, maintaining efficiency while improving temporal dependencies. Additionally, we introduce a selective fine-tuning approach, replacing self-attention layers in SSL models with WSM blocks and fine-tuning only these blocks in low-resource settings. Our approach improves ASR performance while reducing peak VRAM usage by 40\% in the SSL models. WSM blocks have linear-time complexity with enhanced context awareness. Selectively replacing some attention layers reduces compute, memory, and latency, making it ideal for low-resource speech recognition.
title Windowed SummaryMixing: An Efficient Fine-Tuning of Self-Supervised Learning Models for Low-resource Speech Recognition
topic Audio and Speech Processing
Machine Learning
Sound
url https://arxiv.org/abs/2602.09043