19.11. Efficient Speaker Recognition for Mobile Devices

Public examination of a doctoral dissertation in the field of computer science

Doctoral candidate: M.Sc. Evgeny Karpov

Date and venue: 19.11.2011, at 12 noon, Louhela auditorium, Science Park in Joensuu campus 

In this dissertation, Evgeny Karpov has studied novel methods for speaker recognition which are applicable for devices with limited hardware resources like mobile phones. Speaker recognition is a technique to identify speakers from their voices. This technology has been an active topic of research for several decades already and successful applications utilizing it can be found on the market such as remote access to computers and e-banking as well as forensic applications. However, reliable systems require fast and expensive hardware to operate in reasonable time, which is not always the case on mobile phone. This thesis mainly addresses the limitations that are found in typical mobile phone implementation and proposes efficient methods and algorithms to overcome them. 

In this work, state of the art methods in speaker recognition are reviewed first and after that several new methods to improve overall system performance are proposed. Speaker pruning methods presented in this thesis can achieve up to 10 times speed-up in Vector Quantization-based systems and up to 34 times speed-up in Gaussian Mixture-based systems with minor error rate degradation from 16.9 % to 18.5 %. Several common techniques such as voice activity detection and feature vector quantization are analyzed and results of practical experiments are reported. Mobile phone specific design issues, most notably the absence of floating point unit, are discussed and algorithm conversion methods to run on a fixed point processor with as little degradation in accuracy as possible are proposed. Efficient model quantization techniques that have been used in the speech recognition domain but can be easily applied to speaker verification are considered. 

The doctoral dissertation of M.Sc. Evgeny Karpov entitled “Efficient Speaker Recognition for Mobile Devices” will be examined at the Faculty of Science and Forestry. The opponent in the public examination is Professor Zheng-Hua Tan, Aalborg University Denmark and the custos is Professor Pasi Fränti of the University of Eastern Finland. 

Photo available for download at http://www.uef.fi/vaitoskuvat 

Contact: Evgeny Karpov, ekarpov@student.uef.fi, tel. +358 40 031 5284

Publishing year: 2011

Back to this years article listing