雲间烟火The length of the VTL is an important acoustic variation among speakers, where, the length of the vocal tract is different for male and female speakers (i.e. female components of a signal. It is also be useful for whitening the noisy components present in the signal and maintains the frequency attribute of the signal.
Thus, the WLPC can effectively normalize the vocal tract length variations and moreover the frequency warped signal has also improved the perceptual quality and the intelligibility of the signal. The frequency warped signal is then used as an input for the above proposed techniques where the modified features are extracted. Finally, the resultant feature vectors are normalized using Cepstral Mean Normalization (CMN) as described below. VTL is shorter than male VTL). Based on this characteristic, the formant frequencies produced by different speakers can vary up to 25% [21]. It causes a serious mismatch between different speakers due to the change in vocal tract shape [22]. This issue can be reduced by normalizing the VTL. Frequency Warping using LPC
Frequency warping has a high level of significance in improving the performance of the speaker independent speech recognition systems [23]. As speaker independent system involves different speakers, the frequency warping is performed by applying the time varying all pole filter using LPC and it is explained below. Warped Linear Predictive Coding (WLPC)
The standard form of LPC analyzes is employed to divide the signal into a smoothed spectral format. In this research work, warping LPC is applied to change the frequency resolution of an input speech signal. WLPC is an alternative technique of LPC, where the spectral representation of the system is customized using the first order all pass filter. For this purpose, all the unit delay elements are replaced by the 1st order all-pass filters. Applying WLPC has the following benefits.
It facilitates the signal by warping the spectra,
Minimizes the speaker variations and help to preserve the important speech information,
Reduces the bit rate required for a given speech signal, and
Improves the intelligibility and the naturalness of the speech without changing pitch.
Applying WLPC is used to shift the resonant frequencies of the LPC model to the Infinite Impulse Response (IIR) filter, by substituting an all-pole system for each delay element. The LPC warping technique is used to warp an all-pole filter defined by numerator coefficients using a first order all pass substitution with alpha value. It generates a new filter with poles and zeros defined by polynomials thB and A. In this research work, a 12 order LPC autoregressive model is used for warping an input signal and the resultant signal contains an all pole filter coefficients. Next, it resynthesizes the resultant LPC parameters using the noise excitation. The warping parameter α is selected between 0.1 and 0.3. Subsequently, LPC whitening method is adopted to whiten the signal and to preserve the formant frequency
Feature Normalization using
Cepstral
Mean
Normalization
The characteristics of a signal can change according to the microphone distance and the type of microphone used for speech acquisition. CMN is the most widely used to eliminate the speech signal components suffer from channel distortions. It effectively works in removing additive noise as well as channel noise. It uses the simplest method for feature normalization, by forcing the mean value of each element of the Cepstral feature vector, to be zero for all utterances [24].
Generally, the mean value of the signal conveys the spectral characteristics of the microphone and room acoustics, and it is not often reliable in signal processing. The mean value of the Cepstral coefficients is calculated across the whole utterance and then subtracted from each frame. The CMN is performed as follows:
Let X={X0,X1,XT-1} be the Cepstral vectors computed using short term analyzes, then the sample mean is represented by the following expression (1).
1T 1
T Xt (1)
t 0
and, the CMN is represented as follows (2).
X
t Xt (2)
where, signal corresponding to Xt is processed by a linear filter. CMN can help to reduce three types of distortions, namely, environment distortions, channel distortions, and intra speaker effects. Therefore, the robustness can be achieved once the mean vector is subtracted from the feature vectors. The performances of the above proposed techniques are discussed in the next section.
IV. EXPERIMENTAL RESULTS
The performance evaluation of the existing and proposed techniques are presented in this section. Experiments are done with 10 Tamil spoken digits (0-9) and 5 spoken names from 30 different speakers. To make utterance variation, the speakers uttered the same word at different interval of time. The utterances consist of 10 repetitions from 15 male and 15 females. The total size of the dataset is 15*30*10=4500. The utterances were
