We found a match
Your institution may have access to this item. Find your institution then sign in to continue.
- Title
Constant Q Cepstral Coefficients for Automatic Speaker Verification System for Dysarthria Patients.
- Authors
Salim, Shinimol; Ahmad, Waquar
- Abstract
This study aims to develop an automated speaker verification (ASV) system for individuals with dysarthria, a speech disorder affecting their ability to articulate words correctly. The ASV system utilizes a duration modification-based data augmentation technique to create a diverse and robust training set by modifying the duration of the training speech samples. The constant Q cepstral coefficient (CQCC) technique is utilized to extract the features, which is better suited for analyzing speech signals with varying spectral properties. It is helpful in speaker verification systems for speakers with non-uniform frequency distribution, and the CQCC features are insensitive to variations in speech rate and irregular pitch patterns. The study evaluates the performance of the ASV system by comparing CQCC and MFCC features and their combination. The study involved training separate i-vector and x-vector models using MFCC and CQCC features alone and in combination. We also investigated the impact of data augmentation through feature fusion on three levels of dysarthria severity: low, medium, and high. The study results indicated that the proposed system, which combined MFCC and CQCC features with duration modification-based data augmentation, performed significantly better than the baseline system. Specifically, the proposed system improved the i-vector and x-vector model EER (equal error rate) by 15.07% and 22.75%, respectively. The evaluation results show that the ASV system accurately verifies control and dysarthric speakers even with highly impaired speech intelligibility, making it a promising solution for security and authentication applications.
- Subjects
INTELLIGIBILITY of speech; DYSARTHRIA; DATA augmentation; SPEECH disorders; SPEECH; FEATURE extraction
- Publication
Circuits, Systems & Signal Processing, 2024, Vol 43, Issue 2, p1101
- ISSN
0278-081X
- Publication type
Article
- DOI
10.1007/s00034-023-02505-0