Auditory-model based robust feature selection for speech recognition.

Area of Science:

• Auditory Neuroscience
• Speech Processing
• Machine Learning

Background:

• Conventional speech recognition methods often rely on optimizing classification performance.
• Existing dimension-reduction techniques may require labeled training data and struggle with noisy inputs.
• Understanding the human auditory system offers a pathway to more robust feature extraction.

Purpose of the Study:

• To develop a robust dimension-reduction technique for speech recognition features.
• To leverage the inherent robustness of the human auditory system in feature selection.
• To create a method that does not require labeled training data.

Main Methods:

• Utilizing a model of the human auditory system to guide feature selection.
• Selecting features to align the Euclidean geometry of the feature domain with the perceptual domain.
• Employing mel-frequency cepstral coefficients (MFCCs) for recognition experiments.

Main Results:

• The proposed method demonstrates robust dimension-reduction for speech recognition.
• The approach outperforms conventional discriminant-analysis based methods on noisy data.
• Feature selection based on the auditory model does not require labeled training data.
• Selecting MFCCs in their natural order yielded subsets with good performance.

Conclusions:

• A model of the human auditory system provides a robust basis for speech feature dimension-reduction.
• This biologically inspired approach offers advantages over traditional methods, particularly in noisy conditions.
• The method's independence from labeled data makes it broadly applicable.