Exploring different attributes of source information for speaker verification with limited test data.

Area of Science:

• Speech processing
• Biometrics
• Signal analysis

Background:

• Speaker verification systems often struggle with limited test data.
• Traditional features like Mel Frequency Cepstral Coefficients (MFCCs) may not capture all relevant speaker-specific information.

Purpose of the Study:

• To investigate the effectiveness of novel acoustic features for speaker verification under data scarcity.
• To combine multiple source features to improve system performance.

Main Methods:

• Exploration of mel power difference of spectrum in subband.
• Analysis of residual mel frequency cepstral coefficient.
• Application of Discrete Cosine Transform (DCT) on integrated linear prediction residual.
• Feature combination and evaluation on the NIST SRE 2003 database.

Main Results:

• The proposed combination of three source features (mel power difference, residual MFCC, DCT of linear prediction residual) achieved a lower Equal Error Rate (EER) of 20.19% and Decision Cost Function (DCF) of 0.3759.
• This performance surpasses the baseline MFCC feature, which yielded an EER of 22.31% and DCF of 0.4128.
• The individual features capture distinct speaker attributes: periodicity, smoothed spectrum, and glottal signal shape.

Conclusions:

• Combining mel power difference, residual MFCC, and DCT of linear prediction residual offers a robust approach to speaker verification, particularly effective under limited test data conditions.
• These complementary features provide a more comprehensive representation of speaker-specific characteristics than traditional MFCCs alone.