Browsing Elektrik- Elektronik Mühendisliği by Author "The purpose of this thesis is to realize a prosodically guided, syllable based, limited-vocabulary, speaker-independent Turkish word recognizer and studying the effects of various parameters on the recognition rate. Basic recognition units are the syllables which form the words in the vocabulary according to the prosodical rules, of Turkish. The input of the system is the 18-word vocabulary spoken by 4 different speakers. The speech is first filtered with a low-pass filter which has cutoff at 3.5 kHz, then sampled at 8 kHz and fed into the PDP 11/23 microcomputer which processes the data. The output is the best estimate of the word at the input. The endpoints of the syllables are found using pitch-period and energy information. The feature sets used for the test and reference templates consist of the coefficients of a 10-pole LPC filter. The comparison between the test and reference templates is performed by dynamic time warping and log-likelihood. similarity measures. Also Turkish prosodical rules are used for reducing the calculation efforts during the comparison. And finally K-Nearest-Neighbour decision rule gives the best estimate of the word at the input. Various runs with different parameters and different speakers were performed and the observations and results are reported in the thesis."
Browsing Elektrik- Elektronik Mühendisliği by Author
JavaScript is disabled for your browser. Some features of this site may not work without it.
Browsing Elektrik- Elektronik Mühendisliği by Author "The purpose of this thesis is to realize a prosodically guided, syllable based, limited-vocabulary, speaker-independent Turkish word recognizer and studying the effects of various parameters on the recognition rate. Basic recognition units are the syllables which form the words in the vocabulary according to the prosodical rules, of Turkish. The input of the system is the 18-word vocabulary spoken by 4 different speakers. The speech is first filtered with a low-pass filter which has cutoff at 3.5 kHz, then sampled at 8 kHz and fed into the PDP 11/23 microcomputer which processes the data. The output is the best estimate of the word at the input. The endpoints of the syllables are found using pitch-period and energy information. The feature sets used for the test and reference templates consist of the coefficients of a 10-pole LPC filter. The comparison between the test and reference templates is performed by dynamic time warping and log-likelihood. similarity measures. Also Turkish prosodical rules are used for reducing the calculation efforts during the comparison. And finally K-Nearest-Neighbour decision rule gives the best estimate of the word at the input. Various runs with different parameters and different speakers were performed and the observations and results are reported in the thesis."