The BU-MEF System for the Speak & Improve Challenge 2025: Spoken Language Assessment Using Speech and Textual Representations

Abstract

This paper describes the system developed by the BUMEF team for the spoken language assessment task in the Speak & Improve Challenge 2025. The proposed system integrates discrete speech representations with textual embeddings within a language model-based regression framework. Unlike traditional approaches that rely on hand-crafted features or low-level acoustic features, our method extracts word-aligned acoustic embeddings using a large-scale speech encoder, followed by vector quantization to produce interpretable speech tokens. These discrete representations are combined with textual inputs within a BERT-based grader, enabling structured and scalable modeling. Experimental results show that incorporating discrete speech tokens alongside text improves scoring accuracy over text-only baselines, highlighting the effectiveness of the proposed approach. Our submitted system achieved a 3.4% relative improvement over the baseline system on the official evaluation set