Efficient Speech Translation through Model Compression and Knowledge Distillation

TL;DR

This paper presents a system combining model pruning, quantization, and knowledge distillation to significantly reduce the size of speech translation models while maintaining high translation quality.

Contribution

It introduces a novel combination of layer pruning, low-rank adaptation with quantization, and knowledge distillation for efficient speech translation models.

Findings

Up to 50% reduction in model size and storage footprint.

Maintains 97-100% of translation quality of larger models.

Effective for speech translation into German and Chinese.

Abstract

Efficient deployment of large audio-language models for speech translation remains challenging due to their significant computational requirements. In this paper, we address this challenge through our system submissions to the "Model Compression" track at the International Conference on Spoken Language Translation (IWSLT 2025). We experiment with a combination of approaches including iterative layer pruning based on layer importance evaluation, low-rank adaptation with 4-bit quantization (QLoRA), and knowledge distillation. In our experiments, we use Qwen2-Audio-7B-Instruct for speech translation into German and Chinese. Our pruned (student) models achieve up to a 50% reduction in both model parameters and storage footprint, while retaining 97-100% of the translation quality of the in-domain (teacher) models.