Faila Nadhifatul Aryza, Syahroni Hidayat
Speech recognition for individuals with dysarthria remains challenging due to unstable acoustic signals, high temporal variability, and frequent articulatory distortions, all of which hinder the ability of acoustic models to consistently capture phonetic patterns. This study aims to identify the most effective feature extraction strategy among three approaches, namely Wav2Vec 2.0, MFCC combined with Wav2Vec 2.0, and Wavelet-MFCC combined with Wav2Vec 2.0, evaluated using the UA-Speech dataset. All models were trained using the Wav2Vec 2.0 Base architecture with a CTC decoding mechanism to map audio signals to character sequences in an end-to-end manner. The experimental results demonstrate that the MFCC-Wav2Vec 2.0 combination yields the best performance, achieving a Word Error Rate (WER) of 0.2990. These findings indicate that combining traditional acoustic features with self-supervised representations yields a more robust speech recognition system for dysarthric speech. © 2026 by the authors of this article. Published under CC-BY.
Universitas Negeri Semarang, Semarang, Indonesia
