Dual-Bandwidth Spectrogram Analysis for Speaker Verification

• Rafaello Virgilli UFG
• Arnaldo Candido Junior Unesp
• Augusto Seben da Rosa UTFPR
• Frederico S. Oliveira UFG
• Anderson da Silva Soares UFG

Resumo

The variability of the human voice is a challenge for speaker verification systems, influenced by individual traits and environmental conditions. This research introduces a novel approach that uses dual-bandwidth spectrograms with the Fast ResNet-34 neural network architecture for speaker verification. Dual-bandwidth spectrograms are data structures similar to multi-channel images, generated by stacking spectrograms derived from the same audio segment using two different window sizes. In this study, we employed window sizes of 5 ms and 30 ms. This approach captures a wider range of voice features across multiple temporal and spectral resolutions. Our findings demonstrate a statistically significant improvement in system performance, achieving an Equal Error Rate (EER) of 1.64% ±0.13%. This represents a 26% enhancement over the previously reported benchmark EER of 2.22% ±0.05%, validating our hypothesis that dual-bandwidth spectrograms offer a more detailed and comprehensive representation of voice features for accurate speaker verification. Analysis of individual bandwidth contributions reveals that narrowband spectrograms carry more relevant features for speaker verification, while the combination with broadband spectrograms provides complementary information.