Speaker Identification and Verification Using Convolutional Neural Network CNN
Speaker identification and verification are important fields contributing to smart IoT, phone banking, remote login services, E-learning, and other applications. In this work, the speaker identification and verification processes have been experimentally proven to have mutual enhancement effects if...
Saved in:
| Main Authors: | , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Tikrit University
2025-05-01
|
| Series: | Tikrit Journal of Engineering Sciences |
| Subjects: | |
| Online Access: | https://www.tj-es.com/ojs/index.php/tjes/article/view/1746 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Speaker identification and verification are important fields contributing to smart IoT, phone banking, remote login services, E-learning, and other applications. In this work, the speaker identification and verification processes have been experimentally proven to have mutual enhancement effects if they are merged together in a proper manner. Speaker identification and verification work cooperatively so that the verifier will enhance the identifier model. The first step is to identify the speaker using context – independent speech signal, and the identifier model (ID) is trained using a classification model. The model’s outputs are then used to control the verification process as a next step. When the verification result is positive, the first process outcome is approved with high confidence. Otherwise, the negative verification will force the ID process to re-configure itself. The loop continues until both verification and ID agree on the speaker. A multiple Gaussian mixture GMM was used to efficiently model each person’s speech features (MFCC) for using expectation maximization (EM). On the other hand, the conducted experiments showed that the one-dimensional convolutional neural network (1D-CNN) proved its superiority over other models for speaker identification. A novel approach was proposed, proving that little data can be expanded with split-add-noise and train-on-the-fly procedures. In many speaker identification approaches, the specific context was used as a keyword or a password to simplify the processing, requiring big data to achieve high accuracy. It is noteworthy that a small amount of data was enough to efficiently train the proposed model, with a verification error of around 3%, i.e., an accuracy of 97%. Meanwhile, 95% and 96% identification accuracy was achieved using two different datasets. Additionally, the suggested algorithm did not imply using any keyword or password because it is a context-independent approach.
|
|---|---|
| ISSN: | 1813-162X 2312-7589 |