Evaluating Automatic Speaker Recognition systems: An overview of the NIST Speaker Recognition Evaluations (1996-2014)

González-Rodríguez, Joaquín

doi:10.3989/loquens.2014.007

Cited by 28 publications

(14 citation statements)

References 45 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…State-of-the-art SV systems are trained to generally make two decisions, acceptance or rejection. There are two types of decision errors, the FAR referred to accept an impostor speaker (Type I error), and the FRR related to the incorrect rejection of a true speaker (Type II error) [1,2]. These errors compared to a threshold determines the system performance.…”

Section: Introductionmentioning

confidence: 99%

Optimization of False Acceptance/Rejection Rates and Decision Threshold for End-to-End Text-Dependent Speaker Verification Systems

et al. 2019

View full text Add to dashboard Cite

Currently, most Speaker Verification (SV) systems based on neural networks use Cross-Entropy and/or Triplet loss functions. Despite these functions provide competitive results, they might not fully exploit the system performance, because they are not designed to optimize the verification task considering the performance measures, e.g. the Detection Cost Function (DCF) or the Equal Error Rate (EER). This paper proposes a first approach to this issue through the optimization of a loss function based on the DCF. This mechanism allows the end-toend system to directly manage the threshold used to compute the ratio between the False Rejection Rate (FRR) and the False Acceptance Rate (FAR). This way connecting the system training directly to the operating point. Results in a text-dependent speaker verification framework, based on neural network supervectors over the RSR2015 dataset, outperform reference systems using Cross-Entropy and Triplet loss, as well as our previously proposal based on an approximation of the Area Under the Curve (aAUC).

show abstract

Section: Introductionmentioning

confidence: 99%

Optimization of False Acceptance/Rejection Rates and Decision Threshold for End-to-End Text-Dependent Speaker Verification Systems

et al. 2019

View full text Add to dashboard Cite

show abstract

“…Recent work in speaker identification relies on frontend processing that extracts short-time spectral information, usually using the MEL frequency cepstral coefficient (MFCC) features [1]. MFCC is the most prevalent short-term spectral feature despite the fact that it was developed for speakerindependent speech recognition.…”

Section: Introduction and Previous Workmentioning

confidence: 99%

“…It is logical, then, that much speaker-related information is missing from MFCC. Attempts to add voice source information back into speaker identification systems to improve them have met limited success [1], [2], [3], [4], probably due to the difficulty and reliability of estimating the voice source waveform itself. We previously introduced the GLOMM method [5], which was based on detecting glottal events (glottal opening/closing) by detecting times of high linear prediction error.…”

Section: Introduction and Previous Workmentioning

confidence: 99%

Glottal mixture model (GLOMM) for speaker identification on telephone channels

Baggenstoss

Wilkinghoff

Kurth

2017

2017 25th European Signal Processing Conference (EUSIPCO)

View full text Add to dashboard Cite

Abstract-The Glottal Mixture Model (GLOMM) extracts speaker-dependent voice source information from speech data. It has previously been shown to provide speaker identification performance on clean speech comparable to universal background model (UBM), a state of the art method based on MFCC. And, when combined with UBM, the error rate was reduced by a factor of three, showing that the voice source information is largely independent of the information contained in the MFCC, yet holds as much speaker-related information. We now describe how GLOMM can be adapted for telephone quality audio and provide significant error reduction when combined with UBM and I-vector approaches. We demonstrate a factor of two error reduction on the NTIMIT data set with respect to the best published results.

show abstract

“…A sample ROC curve illustrating the trade-off between selectivity and sensitivity of a classifier[30] …”

mentioning

confidence: 99%

Keystroke and Touch-dynamics Based Authentication for Desktop and Mobile Devices

Syed¹

View full text Add to dashboard Cite

The most commonly used system on desktop computers is a simple username and password approach which assumes that only genuine users know their own credentials. Once broken, the system will accept every authentication trial using compromised credentials until the breach is detected. Mobile devices, such as smart phones and tablets, have seen an explosive increase for personal computing and internet browsing. While the primary mode of interaction in such devices is through their touch screen via gestures, the authentication procedures have been inherited from keyboard-based computers, e.g. a Personal Identification Number, or a gesture based password, etc. This work provides contributions to advance two types of behavioral biometrics applicable to desktop and mobile computers: keystroke dynamics and touch dynamics. Keystroke dynamics relies upon the manner of typing rather than what is typed to authenticate users. Similarly, a continual touch based authentication that actively authenticates the user is a more natural alternative for mobile devices. Within the keystroke dynamics domain, habituation refers to the evolution of user typing pattern over time. This work details the significant impact of habituation on user behavior. It offers empirical evidence of the significant impact on authentication systems attempting to identify a genuine user affected by habituation, and the effect of habituation on similarities between users and impostors. It also proposes a novel effective feature for the keystroke dynamics domain called event sequences. We show empirically that unlike features from traditional keystroke dynamics literature, event sequences are independent of typing speed. This provides a unique advantage in distinguishing between users when typing complex text. With respect to touch dynamics, an immense variety of mobile devices are available for consumers, differing in size, aspect ratio, operating systems, hardware and software specifications to name a few. An effective touch based authentication system must be able to work with one user model across a spectrum of devices and user postures. This work uses a locally collected dataset to provide empirical evidence of the significant effect of posture, device size and manufacturer on user authentication performance. Based on the results of this strand of research, we suggest strategies to improve the performance of continual touch based authentication systems. This work is dedicated to my family for the unwavering support, encouragement and understanding. And to my wife, Sameera, my pillar of strength. You are, therefore I am.  To Dr. Bojan Cukic-As an advisor, you have mentored me with patience and dedication. Thank you for helping me build upon my strengths and having faith in me. You went out of your way to provide me opportunities to develop professionally and were easy-going when I needed a break. Your genial nature helped make the Ph.D. so much more bearable and, in many instances, fun. Thank you for being an awesome advisor.

show abstract

Evaluating Automatic Speaker Recognition systems: An overview of the NIST Speaker Recognition Evaluations (1996-2014)

Cited by 28 publications

References 45 publications

Optimization of False Acceptance/Rejection Rates and Decision Threshold for End-to-End Text-Dependent Speaker Verification Systems

Optimization of False Acceptance/Rejection Rates and Decision Threshold for End-to-End Text-Dependent Speaker Verification Systems

Glottal mixture model (GLOMM) for speaker identification on telephone channels

Keystroke and Touch-dynamics Based Authentication for Desktop and Mobile Devices

Contact Info

Product

Resources

About