Kamshat Saduakassova

As a Leverhulme Trust Doctoral Scholar, I explore why automatic speech recognition (ASR) fails for speakers with atypical speech, such as dysarthria or stuttering. The methodology involves analysing ASR outputs on existing speech datasets, using linguistic and acoustic-phonetic analysis to build a taxonomy of error types. This will inform LLM-based post-processing strategies that correct ASR output without hallucinating content or normalising away the speaker’s own way of speaking. Overall, this research project aims to combine linguistic and acoustic-phonetic evidence to make speech-to-text more accessible to the people who rely on it most.