Unfolding A Few Structures for The Many: Memory-Efficient Compression of Conformer and Speech Foundation Models

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Li, Zhaoqing, Xu, Haoning, Xie, Xurong, Jin, Zengrui, Wang, Tianzi, Liu, Xunying
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866916761917456384
author Li, Zhaoqing
Xu, Haoning
Xie, Xurong
Jin, Zengrui
Wang, Tianzi
Liu, Xunying
author_facet Li, Zhaoqing
Xu, Haoning
Xie, Xurong
Jin, Zengrui
Wang, Tianzi
Liu, Xunying
contents This paper presents a novel memory-efficient model compression approach for Conformer ASR and speech foundation systems. Our approach features a unique "small-to-large" design. A compact "seed" model containing a few Conformer or Transformer blocks is trained and unfolded many times to emulate the performance of larger uncompressed models with different logical depths. The seed model and many unfolded paths are jointly trained within a single unfolding cycle. The KL-divergence between the largest unfolded and smallest seed models is used in a self-distillation process to minimize their performance disparity. Experimental results show that our foldable model produces ASR performance comparable to individually constructed Conformer and wav2vec2/HuBERT speech foundation models under various depth configurations, while requiring only minimal memory and storage. Conformer and wav2vec2 models with a reduction of 35% and 30% parameters are obtained without loss of performance, respectively.
format Preprint
id arxiv_https___arxiv_org_abs_2505_21237
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Unfolding A Few Structures for The Many: Memory-Efficient Compression of Conformer and Speech Foundation Models
Li, Zhaoqing
Xu, Haoning
Xie, Xurong
Jin, Zengrui
Wang, Tianzi
Liu, Xunying
Sound
Audio and Speech Processing
This paper presents a novel memory-efficient model compression approach for Conformer ASR and speech foundation systems. Our approach features a unique "small-to-large" design. A compact "seed" model containing a few Conformer or Transformer blocks is trained and unfolded many times to emulate the performance of larger uncompressed models with different logical depths. The seed model and many unfolded paths are jointly trained within a single unfolding cycle. The KL-divergence between the largest unfolded and smallest seed models is used in a self-distillation process to minimize their performance disparity. Experimental results show that our foldable model produces ASR performance comparable to individually constructed Conformer and wav2vec2/HuBERT speech foundation models under various depth configurations, while requiring only minimal memory and storage. Conformer and wav2vec2 models with a reduction of 35% and 30% parameters are obtained without loss of performance, respectively.
title Unfolding A Few Structures for The Many: Memory-Efficient Compression of Conformer and Speech Foundation Models
topic Sound
Audio and Speech Processing
url https://arxiv.org/abs/2505.21237