Tharmarajah Thiruvaran

Group delay features for speaker recognition

Group delay is proposed as an effective means of representing spectral phase information as a feature in speaker recognition. Robustness of group delay features is difficult to achieve, since the spiky nature of the group delay masks the fine structure of the group delay. In this paper, two features based on group delay are proposed by reducing the effect of spikes with two different approaches. The first is log compression, to address the masking effects of the spikes, and the second is to use a sub-band based approach, where masking is restricted within certain bands containing the spikes. The purpose of this paper is to introduce different types of group delay feature extraction methods. The two features are evaluated on the cellular NIST 2001 database.

Evaluation of a fused FM and cepstral-based speaker recognition system on the NIST 2008 SRE

In this paper, the fusion of two speaker recognition subsystems, one based on Frequency Modulation (FM) and another on MFCC features, is reported. The motivation for their fusion was to improve the recognition accuracy across different types of channel variations, since the two features are believed to contain complementary information. It was found that the MFCC-based subsystem outperformed the FM-based subsystem on telephone conversations from NIST SRE-06 dataset, while the opposite was true for NIST SRE-08 telephone data. As a result, the FM-based subsystem performed as well as the MFCC-based subsystem and their fusion gave up to 23% relative improvement in terms of EER over the MFCC subsystem alone, when evaluated on the NIST 2008 core condition.

A study of effectiveness of speech enhancement for cognitive load classification in noisy conditions

In the last decade, speech-features have been effectively utilized for estimating cognitive load level in ideal conditions where recorded speech is clean. However, in more realistic conditions, the recorded speech data is corrupted by noise. Hence, the employment of speech enhancement is essential to reduce the noise. In this paper, the effectiveness of three speech enhancement algorithms proposed in our previous studies are compared based on performance and processing time and the most suitable method is utilized to denoise the input noisy speech before feeding it to a cognitive load classification system in order to improve its performance. The results of this study indicate that the use of speech enhancement can reduce 3.0% of average relative error rate for the system under the effect of various noisy conditions.

Measuring reliability in forensic voice comparison.

Recently there has been a great deal of concern in forensic science about validity and reliability (accuracy and precision). The log‐likelihood‐ratio cost (Cllr), developed for automatic speaker recognition, is increasingly applied as a standard measure of accuracy in forensic voice comparison, but so far there has been little work on developing a metric of precision within this field. Because voice data can have a large amount of intrinsic variation at the source, and likelihood ratios are typically calculated using a single suspect recording and a single offender recording, assessing the precision of a forensic‐voice‐comparison system is extremely important. This presentation discusses the importance of measuring precision and describes two procedures, one parametric and one non‐parametric, for calculating 95% credible intervals for the likelihood ratios resulting from running tests of forensic‐voice‐comparison systems (in which some comparisons are known to be same‐speaker comparisons and others are kn...

Normalization of Modulation Features for Speaker Recognition

features are emerging as a more recent alternative to more conventional magnitude-based features for speech processing applications such as speaker recognition. Frequency modulation (FM) based features are one such example and in this paper their normalization using feature warping is examined in detail. Evaluations of different FM feature and warping configurations on the NIST 2001 Speaker Recognition corpus show that feature warping is not an effective normalization technique for FM features, despite its well-known effectiveness for Mel frequency cepstral coefficients (MFCC). This study further suggests a closer investigation on feature dependency of the existing system when using new features.