Syllable Based Speech Recognition

Abstract

Speech recognition is the process of converting speech signals to the text. Studies on speech recognition have increased very fast for the last twenty-five years. Most of these studies have used phoneme and word as speech recognition units. Namely, in phoneme based speech recognition systems, all phonemes in a language have been modelled by a speech recognition method, and then the phonemes can be detected by these models. The recognized words are constructed with concatenating these phonemes. However, word based systems model the word utterances and try to recognize the word as a text unit. Word based systems have better success rate than phoneme based systems. As a rule, if a speech recognition system has longer recognition unit than sub-word units, it has better success rate on recognition process. In addition, phoneme end-point detection is quite difficult operation, and this effects the success of the system. Turkish, that is one of the least studied languages in the speech recognition field, has different characteristics than European languages which require different language modelling technique. Since Turkish is an agglutinative language, the degree of inflection is very high. So, many words are generated from a Turkish word’s root by adding suffixes. That’s why, word based speech recognition systems are not adequate for large scaled Turkish speech recognition. Because Turkish is a syllabified language and there are approximately 3500 different Turkish syllables, speech recognition system for Turkish will be suitable to use syllable (sub-word unit) as a speech recognition unit. Turkish speech recognition studies have increased in the past decade. These studies were based on self organizing feature map (Artuner, 1994), DTW (Meral, 1996; Ozkan, 1997), HMM (Arisoy & Dutagaci, 2006; Karaca, 1999; Koc, 2002; Salor & Pellom, 2007; Yilmaz, 1999) and Discrete Wavelet Neural Network (DWNN) and Multi-layer Perceptron (MLP) (Avci, 2007). In a simplified way, speech recognizer includes the operations as preprocessing, feature extraction, training, recognition and postprocessing. After the speech recognizer takes the acoustic speech signal as an input, the output of the recognizer will be the recognized text. The most common approaches to speech recognition can be divided into two classes: “template based approach” and “model based approach”. Template based approaches as Dynamic Time Warping (DTW) are the simplest techniques and have the highest accuracy when used properly. The electrical signal from the microphone is digitized by an analog-todigital converter. The system attempts to match the input with a digitized voice sample, or template. The system contains the input template, and attempts to match this template with