Deep neural architectures for dialect classification with single frequency filtering and zero-time windowing feature...

Area of Science:

• Speech Processing
• Machine Learning
• Linguistics

Background:

• Dialect classification is crucial for understanding linguistic variations.
• Traditional methods often struggle with subtle spectro-temporal nuances.
• Single Frequency Filtering (SFF) and Zero-Time Windowing (ZTW) offer enhanced resolution.

Purpose of the Study:

• To investigate advanced signal processing (SFF, ZTW) with Deep Neural Networks (DNNs) for English dialect classification.
• To evaluate novel feature representations against established methods.
• To assess the performance of various DNN architectures (CNN, TCN, TDNN, ECAPA-TDNN).

Main Methods:

• Extracted four feature representations (Spectrogram, Cepstral Coefficients, Mel Filter-Bank Energies, MFCCs) using SFF and ZTW.
• Employed CNN, TCN, TDNN, and ECAPA-TDNN classifiers.
• Conducted experiments with and without data augmentation on the UT-Podcast database.

Main Results:

• Proposed SFF/ZTW features outperformed baseline STFT features by 15-20% without data augmentation.
• TCN, TDNN, and ECAPA-TDNN models further improved performance, especially with SFF-derived features.
• Best accuracy (85.48%) achieved using Single Frequency Filtered Cepstral Coefficients with ECAPA-TDNN.

Conclusions:

• Advanced signal processing techniques (SFF, ZTW) combined with DNNs are highly effective for dialect classification.
• Novel feature representations offer significant improvements over traditional methods.
• The study provides a robust framework and open-source code for future research in dialect identification.