Hybrid machine learning classification scheme for speaker identification.

Area of Science:

• Speech communication
• Machine learning
• Biometrics

Background:

• Next-generation automatic spokesperson recognition (ASR) systems require advanced speaker identification techniques.
• Current methods may not fully leverage combined spectral features for robust speaker characterization.

Purpose of the Study:

• To develop and validate a hybrid machine learning strategy for enhanced speaker identification.
• To fuse multiple spectral features for improved accuracy in ASR systems.

Main Methods:

• Applied a hybrid machine learning strategy combining mel-frequency cepstral coefficients (MFCCs), spectral kurtosis, spectral skewness, normalized pitch frequency (NPF), and formants.
• Utilized a Random Forest-Support Vector Machine (RF-SVM) classifier for speaker identification and verification.
• Validated the approach on standard speaker databases (ELSDSR, TIMIT, NIST) under various noise conditions.

Main Results:

• Achieved an average identification rate of 97% and an equal error rate (EER) below 2% on experimental datasets.
• Demonstrated high accuracy on ELSDSR (98%), TIMIT (99%), and NIST (94%) datasets with low EERs (2%, 1%, 2% respectively).
• The proposed fusion method outperformed individual schemes and other state-of-the-art speaker recognition mechanisms.

Conclusions:

• The hybrid RF-SVM strategy effectively identifies speakers by fusing diverse spectral features.
• This approach offers a robust and superior solution for advanced speaker recognition tasks.
• The method shows significant potential for next-generation ASR systems.