The Concept of Task Specific Speech Database for VAD Systems

Abstract

The aim of this article is to present a concept of small size, low cost, and stile flexible speech database especially designed for the pin-poses of construction and evaluation of voice activity detection systems. As an example of the concept the Slovak speech database for VAD systems is introduced. It covers wide range of SNR ratios that can be automatically adjusted, according to both general and particular applications. It consists of specially selected isolated words with time labeled occurrences which keeps its size small while enabling its automatic handling. The evaluation of the database was performed by 2 VAD algorithms and compared to the SPEECON database (Spanish version). Observed outcomes fully justify its design and encourage its use for the development and evaluation of high performance VAD systems