Fast and Efficient Approaches in Text-dependent Speaker Recognition Systems

Fast and Efficient Approaches in Text-dependent Speaker Recognition Systems

Článek recenzovaný mimo WoS a Scopus

Although text-dependent speaker recognition systems are not so intensively developed as text-independent ones, there are still many applications, where the text dependency is necessary. This is for example a case of biometric systems. With the increasing computational burden and decreasing price of memories it is possible to implement new and sophisticated algorithms that rapidly increase the accuracy of these systems. But there are still some applications, where fast, simple and efficient approaches are necessary. This work introduces new methods of the text-dependent speaker recognition based on voice imprint and classifiers suitable for the processing of this feature vector. It is shown, that using combination of voice imprint and classifier based on fractional distances, it is possible to reach the successful detection rate equal to 96.94 % and 82.29 % in case of microphone mismatch. Speaking about equal error rate, system reached 4.08 % and 10.94 % considering microphone mismatch. It is also shown that voice imprint can approximately 30 times decrease the memory requirements.

Although text-dependent speaker recognition systems are not so intensively developed as text-independent ones, there are still many applications, where the text dependency is necessary. This is for example a case of biometric systems. With the increasing computational burden and decreasing price of memories it is possible to implement new and sophisticated algorithms that rapidly increase the accuracy of these systems. But there are still some applications, where fast, simple and efficient approaches are necessary. This work introduces new methods of the text-dependent speaker recognition based on voice imprint and classifiers suitable for the processing of this feature vector. It is shown, that using combination of voice imprint and classifier based on fractional distances, it is possible to reach the successful detection rate equal to 96.94 % and 82.29 % in case of microphone mismatch. Speaking about equal error rate, system reached 4.08 % and 10.94 % considering microphone mismatch. It is also shown that voice imprint can approximately 30 times decrease the memory requirements.

text-dependent speaker recognition, multi-section vector quantization

text-dependent speaker recognition, multi-section vector quantization

MEKYSKA, J.; FAÚNDEZ ZANUY, M.; FABREGAS, J.; SMÉKAL, Z.

2012

01.04.2011

1213-1539

Elektrorevue - Internetový časopis (http://www.elektrorevue.cz)

2

1

Česká republika

1

6

6