Robust text-independent speaker identification and verification using multi-feature fusion and student’s t modelling

Abstrakt

This paper presents a text-independent speaker
identification system that utilizes MFCC, LPC, prosody, and
optimized multi-level DWT features for robust speaker modeling.
The system is designed for multiple standard speech databases,
including TIMIT, NTIMIT, SITW, and NIST2008. During training,
features from each speaker are normalized to zero-mean
and unit-variance, and Student-t distributions are fitted to model
the statistical characteristics of each speaker. For testing, features
are normalized using the corresponding speaker’s training
statistics, and speaker identity is predicted based on maximum
log-likelihood estimation over the trained models. Experimental
results confirm the superiority of the proposed system, which
achieves high accuracy across multiple datasets (e.g., 98.33% on
TIMIT, 89.38% on NTIMIT, 96.88% on SITW, and 100.00%
on NIST2008) and consistently outperforms existing state-of-theart
methods under AWGN conditions, demonstrating significant
improvements in identification accuracy and the effectiveness of
multi-feature fusion and Student-t modeling.