Noise-robust text-dependent speaker identification using cochlear models.

Area of Science:

• Acoustics and Signal Processing
• Bioacoustics
• Computational Auditory Neuroscience

Background:

• Speaker identification (SID) systems struggle with noise-robustness.
• Human auditory systems exhibit remarkable noise-robustness in speaker identification.
• Conventional SID approaches often lack the biological plausibility for optimal noise handling.

Purpose of the Study:

• To develop a noise-robust text-dependent speaker identification system.
• To evaluate the performance of a novel cochlear model (CARFAC) for speaker identification in noisy conditions.
• To compare CARFAC's efficacy against established auditory feature extraction methods.

Main Methods:

• Implementation of a real-time cochlear model, CARFAC (cascade of asymmetric resonators with fast-acting compression).
• Testing the CARFAC-based SID system on noisy speech signals across various noise types and levels.
• Comparative analysis with mel-frequency cepstrum coefficients (MFCCs), frequency domain linear prediction (FDLP), and an auditory nerve model.

Main Results:

• The CARFAC-based system significantly outperformed conventional methods in noise-robust speaker identification.
• Consistent superior performance was observed across diverse datasets, noise conditions, speaking rates, and classifiers.
• CARFAC's enhanced noise-robustness is attributed to its nonlinear processing of auditory input signals.

Conclusions:

• The CARFAC model offers a promising bio-inspired approach for developing advanced, noise-robust speaker identification systems.
• The inherent nonlinearities within the CARFAC model are crucial for achieving human-like noise-robustness in speaker recognition.
• This research suggests that mimicking the nonlinear processing of the human auditory system is key to overcoming noise challenges in SID.