Paper Info
Title
Gaussian Filter Based Data-Driven Cepstral Features for Robust Speaker Verification System
Abstract
In different speech-based applications, the speech-signal-based frequency cepstral coefficient (SFCC) based feature extraction method has been used successfully. The conventional method for extraction of such feature uses triangular filter banks similar to mel frequency cepstral coefficient (MFCC) features. In this work, we first present Gaussian filter based speech-signal-based frequency cepstral coefficient (GSFCC) based feature extraction technique derived from the speech-signal-based scale. Next, we use spectral entropy modulated speech-signal-based scale along with Gaussian filter to derive new cepstral feature for ASV task. We find improved performance of proposed features over baseline method in both clean and noisy conditions. The experiments were carried on NIST SRE 2001 database with simulated noise. In addition, a more recent real-world, noisy database VoxCeleb1 is used in the study. Finally, we do score level fusion of auditory filter based cepstral feature with data-driven filter based cepstral features. The performance obtained by the proposed features gives up to 15.71% and 15.37% relative improvement in equal error rate (EER) over the baseline method in NIST SRE and VoxCeleb1 databases, respectively.
Year
DOI
Venue
2019
10.1109/ICIT48102.2019.00027
2019 International Conference on Information Technology (ICIT)
Keywords
DocType
ISBN
Speaker vrification,MFCC,mel scale,Speech signal-based scale,SFCC,Noisy condition,i-vector,Spectral entropy,Gaussian filter
Conference
978-1-7281-6053-5
Citations 
PageRank 
References 
0
0.34
13
Authors
2
Name
Order
Citations
PageRank
Susanta Kumar Sarangi100.34
Goutam Saha255.15