Efficient Acoustic Front-End Processing for Tamil ...(IJIGSP-V8-N7-3)(9)

Parametric techniques are better than non-parametric techniques in performing PSD estimation. In the previous experiments, the performance evaluation of windowing methods and PSD estimation is done [20]. The performance comparison of Yule Walker AR and Welch methods are evaluated based on both subjective and objective measures. It is noticed from the experimental outcomes that, the Yule walker AR method has produced better results. Based on the outcome, the Yule-Walker method is implemented to estimate the PSD of an input speech signal. It fits an Auto Regressive (AR) model to the windowed input data so as to minimize the forward prediction error in the least squares sense. This is done by Levinson-Durbin recursion algorithm. The output column vector contains the estimate of the PSD with equally spaced frequency points.

Multi Taper Windowing for Yule Walker AR Power Spectrum

The particular novelty of the proposed method is, instead of performing a regular feature extraction from an input signal, the metrics of multi taper windowing, power spectrum and weight estimation are additionally given as an input to the feature extraction. It is implemented as follows:-

The pre-processed signal which is divided into short frames are multiplied by the multi taper windowing method,

The spectrum and the multi taper weight have been estimated,

Subsequently, Yule walker AR power spectrum is estimated using Levinson-Durbin Recursion algorithm, and

In the final stage, mean value of a spectrum, multi taper weight and the estimated power spectrum from the Yule Walker AR method has been added to the signal.

The detailed descriptions of the steps involved in MTYW-GFCC feature extraction are presented in the following algorithm. In this research work, better performance has been achieved when the frame size is assigned to 240, frame shift is assigned to th120 and the number of tapers is assigned to 8. By using 4 order Yule Walker AR spectrographic analyzes, an important region of signal which is related to the speech signal is determined.

Efficient Acoustic Front-End Processing for Tamil Speech Recognition using Modified GFCC Features

The experimental results confirm that the MTYW-GFCC features have increased the recognition accuracy and it is presented in section 4. Based on the significant improvements achieved by the above feature extraction techniques, the combinational features are implemented subsequently.

Proposed Algorithm

roots generated by the LPC that represent the vocal tract filter.

Using the normalized FFT, the amplitude of the formant frequencies is obtained. Figure 5 shows the first five formant frequencies extracted for the word ―poojiam‖. Since these features are specifically used to distinguish vowels, it helps to recognize Tamil spoken words depend on the vowels present in an input signal. Better results are achieved with these combinational features, since the formant frequencies have the ability to reduce the mismatch between training and deployment environment.

Fig.5. First Five Formants Extracted for the Word ―poojiam‖

2) Combinational Features using MTYW-GFCC with Formant Frequencies (MTYW-GFCC-FF)

In order to make discrimination between each word, the combinational features are proposed by combining the MTYW-GFCC features with Formant Frequencies. Formant frequencies are vocal tract resonances, the frequency of which depends on the length of the tongue or tongue advancement. It can effectively distinguish the meaningful frequency components of human speech patterns. The first two formants are particularly important in speech recognition. At least three formants are generally required and up to five formants are needed for achieving high performance. In this work, the first five formant frequencies are obtained from the polynomial

3) Frequency Warping and Feature Normalization using LPC and CMN (FWCMN-MTYW-GFCC-FF)

Next, an attempt is made to reduce the speaker and channel variations which affect the ASR performance. For this purpose, frequency warping and feature normalization techniques are implemented after applying the above proposed combinational features. In this research work, initially, the frequency warping is applied to the five pass pre-processed input signal and then the modified features and combinational features are extracted. Finally, the combinational feature vectors are normalized by applying the CMN technique. Figure 6 shows the frequency warping and feature normalization using LPC and CMN.

Fig.6. Frequency Warping and Feature Normalization using LPC and CMN

搜索“diyifanwen.net”或“第一范文网”即可找到本站免费阅读全部范文。收藏本站方便下次阅读，第一范文网，提供最新初中教育Efficient Acoustic Front-End Processing for Tamil ...(IJIGSP-V8-N7-3)(9)全文阅读和word下载服务。