Title

The Effect of Using FFT and Bandstop Filter on Voice Recordings with Different Amounts of Background Noise on the Accuracy of a Speaker Verification System

Author(s)

Lisa C. Qu

School Name

Spring Valley High School

Grade Level

10th Grade

Presentation Topic

Math and Computer Science

Presentation Type

Non-Mentored

Oral Presentation Award

1st Place

Written Paper Award

1st Place

Abstract

Voice recognition systems are often used in environments with background noise, which must be removed to improve their accuracy. The purpose was to develop a speaker verification system using fast Fourier transform (FFT) and bandstop filtering to more accurately recognize voices with background noise. One hypothesis was that if FFT and bandstop filter were used to remove background noise from voices, then these voices would be more accurately recognized by the system than ones that were not transformed. It was also hypothesized that if the false acceptance rate (FAR) of the system increased, then the false rejection rate (FRR) would decrease. MATLAB was used to program the mel-frequency cepstral coefficients (MFCC) speaker verification system. Control and background noise voices were tested both using and not using FFT and bandstop filtering. After filtering, the accuracy of the system for the control voices decreased from 100% to 0% and remained constant at 0% after filtering for background noise voices. A paired t-test was conducted on the accuracies of the system before and after filtering, and t(1) = 1.00, p > 0.05. The hypothesis that the accuracy of the system would increase after FFT and bandstop filtering was not supported. Paired sign tests were conducted on the differences between the FRRs and FARs of the voices before and after filtering. For the difference between the non-filtered and filtered control FRRs, p = 1.00. For the difference between the non-filtered and filtered background noise FARs, the test value was 0.00, p < 0.05. Also, for the difference between the non-filtered and filtered background noise FRRs, p = 1.00. The hypothesis that using FFT and a bandstop filter would decrease the FARs and the FRRs of the system was not supported. Correlation was conducted on the FAR and FRR of the non-filtered background noise voices, and r(5) = 0.931, p < 0.05 for the FAR, and r(5) = -0.931, p < 0.05 for the FRR. The second hypothesis was supported because as the FAR increased, the FRR decreased.

Start Date

4-11-2015 2:00 PM

End Date

4-11-2015 2:15 PM

COinS
 
Apr 11th, 2:00 PM Apr 11th, 2:15 PM

The Effect of Using FFT and Bandstop Filter on Voice Recordings with Different Amounts of Background Noise on the Accuracy of a Speaker Verification System

Voice recognition systems are often used in environments with background noise, which must be removed to improve their accuracy. The purpose was to develop a speaker verification system using fast Fourier transform (FFT) and bandstop filtering to more accurately recognize voices with background noise. One hypothesis was that if FFT and bandstop filter were used to remove background noise from voices, then these voices would be more accurately recognized by the system than ones that were not transformed. It was also hypothesized that if the false acceptance rate (FAR) of the system increased, then the false rejection rate (FRR) would decrease. MATLAB was used to program the mel-frequency cepstral coefficients (MFCC) speaker verification system. Control and background noise voices were tested both using and not using FFT and bandstop filtering. After filtering, the accuracy of the system for the control voices decreased from 100% to 0% and remained constant at 0% after filtering for background noise voices. A paired t-test was conducted on the accuracies of the system before and after filtering, and t(1) = 1.00, p > 0.05. The hypothesis that the accuracy of the system would increase after FFT and bandstop filtering was not supported. Paired sign tests were conducted on the differences between the FRRs and FARs of the voices before and after filtering. For the difference between the non-filtered and filtered control FRRs, p = 1.00. For the difference between the non-filtered and filtered background noise FARs, the test value was 0.00, p < 0.05. Also, for the difference between the non-filtered and filtered background noise FRRs, p = 1.00. The hypothesis that using FFT and a bandstop filter would decrease the FARs and the FRRs of the system was not supported. Correlation was conducted on the FAR and FRR of the non-filtered background noise voices, and r(5) = 0.931, p < 0.05 for the FAR, and r(5) = -0.931, p < 0.05 for the FRR. The second hypothesis was supported because as the FAR increased, the FRR decreased.