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

TL;DR

This paper introduces Llama-SMoP, a scalable, efficient multimodal LLM for AVSR that maintains high performance in noisy environments by using a Sparse Mixture of Projectors to reduce computational costs.

Contribution

The paper presents Llama-SMoP, a novel multimodal LLM architecture employing sparse MoE projectors, enabling high-performance AVSR with lower inference costs and improved noise robustness.

Findings

Llama-SMoP DEDR outperforms other configurations on AVSR tasks.

Sparse MoE projectors maintain performance with smaller models.

Ablation studies validate effectiveness in noise robustness.

Abstract

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.