Related Experiment Video
Updated: Jan 10, 2026
06:04
Systematic Hearing Performance Evaluation Process for Adolescents with Cochlear Implantation at Early Ages
Published on: March 24, 2023
736
Classification of Speech and Associated EEG Responses from Normal-Hearing and Cochlear Implant Talkers Using Support
Shruthi Raghavendra1, Sungmin Lee2, Chin-Tuan Tan3
1Independent Researcher, Santa Clarita, CA 91350, USA.
Audiology Research
|November 24, 2025
Summary
Support Vector Machines effectively distinguish speech from cochlear implant (CI) users and normal-hearing (NH) individuals using acoustic and electroencephalogram (EEG) features. Gaussian and RBF kernels showed superior performance, aiding hearing research and rehabilitation.
Area of Science:
- Auditory Neuroscience
- Speech Processing
- Machine Learning in Healthcare
Background:
- Speech intelligibility in cochlear implant (CI) users varies significantly, leading to distinct acoustic characteristics compared to normal-hearing (NH) individuals.
- Existing CI technology supports speech acquisition but has limitations in achieving consistent intelligibility.
- Understanding these differences is crucial for advancing auditory rehabilitation and speech processing technologies.
Purpose of the Study:
- To evaluate the efficacy of Support Vector Machine (SVM) algorithms in classifying speech from NH and CI talkers.
- To compare the performance of different SVM kernel functions (Linear, Polynomial, Gaussian, RBF) using acoustic and EEG features.
- To assess the potential of both speech-derived and EEG-derived features for differentiating between NH and CI speech.
Main Methods:
- Speech data were collected from 8 CI and 8 NH talkers.
- Electroencephalogram (EEG) responses were recorded from 11 NH listeners exposed to the speech stimuli.
- Six acoustic features (Energy, ZCR, Pitch, LPC, MFCCs, PLP-CC) were extracted from speech and EEG signals.
- SVM classifiers with 3-fold cross-validation were used to assess classification accuracy.
Main Results:
- SVMs achieved 100% and 94% classification accuracy for Energy and MFCC features, respectively, using Gaussian and RBF kernels on speech signals.
- EEG analysis using SVMs yielded >70% accuracy for ZCR and Pitch features with Gaussian and RBF kernels.
- Gaussian and RBF kernels demonstrated superior performance, particularly with Energy and MFCC features.
Conclusions:
- Both speech-derived and EEG-derived features effectively differentiate between CI and NH talkers.
- SVMs, particularly with Gaussian and RBF kernels, are highly effective for this classification task.
- Findings support the use of SVMs in multimodal hearing research for improving CI speech processing and auditory rehabilitation.
