Scaling and Enhancing LLM-based AVSR: A Sparse Mixture of Projectors Approach

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Cappellazzo, Umberto, Kim, Minsu, Petridis, Stavros, Falavigna, Daniele, Brutti, Alessio
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866915295988285440
author Cappellazzo, Umberto
Kim, Minsu
Petridis, Stavros
Falavigna, Daniele
Brutti, Alessio
author_facet Cappellazzo, Umberto
Kim, Minsu
Petridis, Stavros
Falavigna, Daniele
Brutti, Alessio
contents Audio-Visual Speech Recognition (AVSR) enhances robustness in noisy environments by integrating visual cues. While recent advances integrate Large Language Models (LLMs) into AVSR, their high computational cost hinders deployment in resource-constrained settings. To address this, we propose Llama-SMoP, an efficient Multimodal LLM that employs a Sparse Mixture of Projectors (SMoP) module to scale model capacity without increasing inference costs. By incorporating sparsely-gated mixture-of-experts (MoE) projectors, Llama-SMoP enables the use of smaller LLMs while maintaining strong performance. We explore three SMoP configurations and show that Llama-SMoP DEDR (Disjoint-Experts, Disjoint-Routers), which uses modality-specific routers and experts, achieves superior performance on ASR, VSR, and AVSR tasks. Ablation studies confirm its effectiveness in expert activation, scalability, and noise robustness.
format Preprint
id arxiv_https___arxiv_org_abs_2505_14336
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Scaling and Enhancing LLM-based AVSR: A Sparse Mixture of Projectors Approach
Cappellazzo, Umberto
Kim, Minsu
Petridis, Stavros
Falavigna, Daniele
Brutti, Alessio
Audio and Speech Processing
Computer Vision and Pattern Recognition
Multimedia
Sound
Audio-Visual Speech Recognition (AVSR) enhances robustness in noisy environments by integrating visual cues. While recent advances integrate Large Language Models (LLMs) into AVSR, their high computational cost hinders deployment in resource-constrained settings. To address this, we propose Llama-SMoP, an efficient Multimodal LLM that employs a Sparse Mixture of Projectors (SMoP) module to scale model capacity without increasing inference costs. By incorporating sparsely-gated mixture-of-experts (MoE) projectors, Llama-SMoP enables the use of smaller LLMs while maintaining strong performance. We explore three SMoP configurations and show that Llama-SMoP DEDR (Disjoint-Experts, Disjoint-Routers), which uses modality-specific routers and experts, achieves superior performance on ASR, VSR, and AVSR tasks. Ablation studies confirm its effectiveness in expert activation, scalability, and noise robustness.
title Scaling and Enhancing LLM-based AVSR: A Sparse Mixture of Projectors Approach
topic Audio and Speech Processing
Computer Vision and Pattern Recognition
Multimedia
Sound
url https://arxiv.org/abs/2505.14336