Closed-set automatic speaker identification using multi-scale recurrent networks in non-native children.

Area of Science:

• Speech processing
• Biometrics
• Machine learning

Background:

• Automatic speaker identification (ASI) has applications in child security, safety, and education.
• Existing ASI systems often struggle with non-native speakers and children due to fluency variations.
• Mel frequency cepstral coefficients (MFCCs) are common but can lose high-frequency information.

Purpose of the Study:

• To develop a closed-set child speaker identification system for non-native English speakers.
• To evaluate the system's performance in both text-dependent and text-independent speech tasks.
• To analyze the impact of speaker fluency on identification accuracy.

Main Methods:

• Utilized multi-scale wavelet scattering transform to preserve high-frequency information, overcoming MFCC limitations.
• Employed a wavelet scattered Bidirectional Long Short-Term Memory (Bi-LSTM) network for speaker identification.
• Assessed system performance using accuracy, precision, recall, and F-measure metrics.

Main Results:

• The proposed wavelet scattered Bi-LSTM system achieved high performance in identifying non-native child speakers.
• The system demonstrated effectiveness in both text-dependent and text-independent speech tasks.
• Performance metrics indicated superiority over existing child speaker identification models.

Conclusions:

• The developed ASI system is effective for identifying non-native child speakers across different speech tasks.
• The multi-scale wavelet scattering transform and Bi-LSTM architecture enhance robustness to fluency variations.
• This technology holds promise for improving child-focused security, safety, and educational applications.