MaLa-ASR: Multimedia-Assisted LLM-Based ASR

TL;DR

MaLa-ASR is a novel LLM-based automatic speech recognition model that leverages multimedia auxiliary information, such as presentation slide keywords, to significantly improve recognition accuracy on conference speech datasets.

Contribution

This paper introduces MaLa-ASR, the first LLM-based ASR model that effectively integrates textual auxiliary data to enhance speech recognition performance.

Findings

Achieves average WERs of 9.4% and 11.7% on SlideSpeech subsets.

Reduces biased word error rate (B-WER) by over 44%.

Sets new state-of-the-art results on the dataset.

Abstract

As more and more information-rich data like video become available, utilizing multi-modal auxiliary information to enhance audio tasks has sparked widespread research interest. The recent surge in research on LLM-based audio models provides fresh perspectives for tackling audio tasks. Given that LLM can flexibly ingest multiple inputs, we propose MaLa-ASR, an LLM-based ASR model that can integrate textual keywords extracted from presentation slides to improve recognition of conference content. MaLa-ASR yields average WERs of 9.4% and 11.7% on the L95 and S95 subsets of the SlideSpeech corpus, representing a significant relative WER drop of 27.9% and 44.7% over the baseline model reported in SlideSpeech. MaLa-ASR underscores LLM's strong performance in speech tasks and the capability to integrate auxiliary information conveniently. By adding keywords to the input prompt, the biased word error rate (B-WER) reduces relatively by 46.0% and 44.2%, establishing a new SOTA on this dataset.