Multiple Feature Fusion for Speaker Recognition from Limited-bandwidth Speech
2
조회
0
다운로드
0
인용
0
참고문헌
키워드
초록
Speaker recognition technology identifies individuals independently of the speech content based on their unique voice characteristics, such as timbre, pitch, formants, and prosody. These vocal traits are leveraged to reliably authenticate or differentiate between speakers in various applications, offering a robust approach to secure and efficient identity verification. This technology aims to determine whether a voice belongs to a registered speaker, especially in cases where fraudulent activities are common, such as voice phishing. Typically, speaker recognition models are trained on datasets such as VoxCeleb, which are sampled at 16 kHz. However, phone communication often involves a lower sampling rate, specifically 8 kHz. To ensure robust speaker recognition in these environments, it is necessary to develop models that can function effectively even with limited-bandwidth speech data. In this study, we aim to mitigate the degradation in the performance of speaker recognition systems for limited-bandwidth speech by extracting and combining various speech features. Specifically, we extract multiple spectrogram forms (vanilla, mel, linear, and MFCC), as well as features such as the constant Q-transform (CQT) and the CCTZ set, which includes chroma, contrast, tonnetz, and the zero-crossing rate. These features are fused in various configurations to enhance the robustness of the model. The experimental results reveal that the fusion of multiple features outperforms the use of single features alone. Moreover, we observed an approximate 0.65% improvement in the equal error rate (EER) when the model trained on 16 kHz data was tested on 8 kHz speech compared to its performance without such feature combinations. These findings highlight the effectiveness of feature fusion for enhancing speaker recognition for limited-bandwidth speech in real-world telecommunication environments.
Abstract
Speaker recognition technology identifies individuals independently of the speech content based on their unique voice characteristics, such as timbre, pitch, formants, and prosody. These vocal traits are leveraged to reliably authenticate or differentiate between speakers in various applications, offering a robust approach to secure and efficient identity verification. This technology aims to determine whether a voice belongs to a registered speaker, especially in cases where fraudulent activities are common, such as voice phishing. Typically, speaker recognition models are trained on datasets such as VoxCeleb, which are sampled at 16 kHz. However, phone communication often involves a lower sampling rate, specifically 8 kHz. To ensure robust speaker recognition in these environments, it is necessary to develop models that can function effectively even with limited-bandwidth speech data. In this study, we aim to mitigate the degradation in the performance of speaker recognition systems for limited-bandwidth speech by extracting and combining various speech features. Specifically, we extract multiple spectrogram forms (vanilla, mel, linear, and MFCC), as well as features such as the constant Q-transform (CQT) and the CCTZ set, which includes chroma, contrast, tonnetz, and the zero-crossing rate. These features are fused in various configurations to enhance the robustness of the model. The experimental results reveal that the fusion of multiple features outperforms the use of single features alone. Moreover, we observed an approximate 0.65% improvement in the equal error rate (EER) when the model trained on 16 kHz data was tested on 8 kHz speech compared to its performance without such feature combinations. These findings highlight the effectiveness of feature fusion for enhancing speaker recognition for limited-bandwidth speech in real-world telecommunication environments.