A Robust Speaker Identification System Using the Responses from a Model of the Auditory Periphery

Area of Science:

• Speech Processing
• Computational Neuroscience
• Biomedical Engineering

Background:

• Speaker identification is challenging in noisy environments.
• Neural responses exhibit inherent robustness to noise.
• Existing methods struggle with noise interference.

Purpose of the Study:

• To propose a novel speaker identification system leveraging neural responses.
• To evaluate the system's performance and robustness under noisy conditions.
• To compare the proposed method against traditional speaker identification techniques.

Main Methods:

• Simulated auditory-nerve fiber responses to speech signals to create 2-D neurograms.
• Trained neurogram coefficients using Gaussian mixture model-universal background model (GMM-UBM).
• Tested on multiple text-dependent and independent speaker databases with various noise types and SNRs.

Main Results:

• The neurogram-based system achieved comparable accuracy to traditional methods in quiet conditions.
• The proposed method significantly reduced classification error rates in noisy environments.
• Demonstrated robustness against white Gaussian, pink, and street noises.

Conclusions:

• Neural response-based neurograms offer a robust feature for speaker identification.
• The proposed system outperforms traditional methods in noisy conditions.
• This approach holds promise for improving speech processing applications in real-world scenarios.