A high-performance vowel spotting system based on a multistage architecture

In this paper we present a novel multi-level vowel detection system with improved accuracy. Multi layer perceptrons (MLP), Discrete Hidden Markov Models (DSHMM) and heuristic rules are combined in three different levels to reduce the probability of false acceptance and rejection of vowel sounds. The TIMIT database was used to train and test this system. The rules are variable and are automatically customized by statistics extracted from the database, which concern the duration, the energy and the distance between vowels. The proposed method can easily be extended to languages other than English as long as a proper database exists for training the system. Its accuracy was measured to 99.22% using all the test data sets of the TIMIT database. Thus, the proposed vowel detection process can be reliably used for speech processing applications (speaker or speech recognition) where accurate vowel spotting algorithms are necessary.