MMSD-Net: Towards Multi-modal Stuttering Detection

TL;DR

MMSD-Net addresses the challenge of detecting stuttering by leveraging multi-modal signals—audio, visual facial cues, and language—through a tripartite transformer encoder architecture combined with a multimodal fusion mechanism and a pretrained Multimodal Language Model. The method compresses modality-specific features, fuses them with attention, and performs binary stuttering classification using a cross-modal decoder. Empirical results on public datasets show MMSD-Net outperforms state-of-the-art uni-modal baselines with a 2%–17% gain in F1-score, demonstrating the added value of visual information for stuttering detection. This work advances practical speech processing systems by enabling context-aware detection of stuttering through integrated audio-visual signals.

Abstract

Stuttering is a common speech impediment that is caused by irregular disruptions in speech production, affecting over 70 million people across the world. Standard automatic speech processing tools do not take speech ailments into account and are thereby not able to generate meaningful results when presented with stuttered speech as input. The automatic detection of stuttering is an integral step towards building efficient, context-aware speech processing systems. While previous approaches explore both statistical and neural approaches for stuttering detection, all of these methods are uni-modal in nature. This paper presents MMSD-Net, the first multi-modal neural framework for stuttering detection. Experiments and results demonstrate that incorporating the visual signal significantly aids stuttering detection, and our model yields an improvement of 2-17% in the F1-score over existing state-of-the-art uni-modal approaches.

Figures (2)

• Figure 1: Illustration of multi-modal cues during stuttering. Sentence: "I was working on fluency shaping techniques."
• Figure 2: Detailed architecture of our proposed model MMSD-Net