Measuring the Redundancy of Decoder Layers in SpeechLLMs

TL;DR

This paper investigates the redundancy in decoder layers of SpeechLLMs, demonstrating that a significant portion of decoder capacity can be pruned without sacrificing performance across speech tasks and languages.

Contribution

It reveals that decoder redundancy is inherited from pretrained LLMs and identifies a global redundancy structure enabling effective pruning for multiple speech tasks.

Findings

7-8B models retain performance with only 60% of decoder layers

Redundant blocks are consistent across different tasks and languages

A single pruned backbone can support multiple speech tasks

Abstract

Speech Large Language Models route speech encoder representations into an LLM decoder that typically accounts for over 90% of total parameters. We study how much of this decoder capacity is actually needed for speech tasks. Across two LLM families and three scales (1-8B), we show that decoder redundancy is largely inherited from the pretrained LLM: text and speech inputs yield similar redundant blocks. We then measure excess capacity by pruning decoder layers and analysing post-pruning healing to increase robustness. Our findings show that 7-8B models retain good ASR performance with only 60% of decoder layers, and the same trend extends to smaller scales with reduced pruning tolerance. We then generalise to speech translation, and show that the same blocks of layers are redundant across speech encoders, tasks and languages, indicating that a more global redundancy structure exists, enabling a single pruned and multi-tasks SpeechLLM backbone to be deployed.