Large vocabulary continuous speech recognition using hidden Markov model

In computer sciences researches on speech can be viewed in three major fields, which are speech recognition, speaker recognition-verification and speech synthesis. The aim of speech recognition, as a major field, is information extraction from speech data by machine or computer. With this thesis an application for large vocabulary continuous speech recognition requirements, which is a portion of multi dimensional speech recognition problem space, is built. On this purpose mel frequency cepstral coefficients (MFCC) feature extraction is used and feature vectors are classified by hidden Markov models (HMM). MFCC is an effective feature extraction method for speaker independent recognition and HMM can handle sequential feature vectors. Real valued MFCC feature vectors are firstly clustered into discrete observations with KMeans algorithm. Discrete observations are then used with left to right, discrete probability emitting HMM models. Statistical nature of speech data can be figured out by the HMM model with no need to labeling, segmentation or signing word start and endings. As HMM training is made on phonemes, test speech data can be decoded with flexibly built test HMM model. System is tested on several models such as, two word discrete recognition, nine words discrete recognition, repetition constraint continuous recognition. Keywords: Large vocabulary continuous speech recognition, hidden Markov model, mel frequency cepstral coefficients.