Effectiveness of Speech-LLMs for Users with Inherently Disfluent Speech

Determine whether users with inherently disfluent speech can interact effectively with Speech Large Language Models (Speech-LLMs).

Background

The paper argues that prior benchmarks emphasize linguistic competence or environmental noise but devote less attention to speech-internal phenomena such as phoneme structure, prosodic patterns, and overlapping speech. Because of this gap, it is uncertain whether individuals whose speech is inherently disfluent can effectively use Speech-LLMs.

To address this, the authors introduce VocalBench-DF, a benchmark designed to systematically evaluate Speech-LLMs under a wide range of disfluency conditions. The open question motivates the need for rigorous evaluation and methods that ensure accessibility for users with disfluent speech.