Highlighting interlanguage phoneme differences based on similarity matrices and convolutional neural network.

Area of Science:

• Linguistics
• Computational Linguistics
• Speech Science

Background:

• Distinguishing between phonemes of different languages is crucial for linguistic analysis.
• Acoustic properties of speech sounds offer a quantifiable basis for comparison.
• Convolutional Neural Networks (CNNs) have shown promise in analyzing complex data like speech spectrograms.

Purpose of the Study:

• To develop and evaluate a method for highlighting acoustic differences between Polish and Lithuanian consonant phonemes.
• To compare the effectiveness of similarity matrices with other established methods for phoneme discrimination.

Main Methods:

• Employing similarity matrices derived from speech acoustic parameters.
• Utilizing convolutional neural networks (CNNs) in conjunction with similarity matrices.
• Comparing results from similarity matrices (with and without parameter reduction) against CNNs with spectrograms, acoustic parameter vectors, k-nearest neighbors, and support vector machines.

Main Results:

• Similarity matrices built on an extensive set of acoustic parameters achieved higher accuracy than those with reduced, non-correlated parameters.
• The similarity matrix approach demonstrated superior performance compared to spectrograms with CNNs, acoustic parameter vectors, k-nearest neighbors, and support vector machines.

Conclusions:

• Similarity matrices based on acoustic parameters are highly effective for discerning consonant phoneme differences between languages.
• The proposed similarity matrix method offers a more accurate approach to phoneme discrimination than current comparative techniques.