雲间烟火4) Tempo Adjustment
Speaking rate is a significant factor in speech recognition applications. Therefore, speaking in a faster or slower manner also has influence on the speech signal. However, even normal speakers will have a tendency to speak faster when using a speech recognition system. But, speaking rate affects both temporal and spectral characteristics of the signal. Therefore, the performance of the acoustic model will also degrade. Benzeghiba, M (2007) says that, the faster speaking rates may also result in more frequent and stronger pronunciation changes [16]. Likewise, Matthew Richardson et al. (1999) say that, the state-of-the-art of an ASR system perform significantly worse on fast speech [17]. Therefore, in this research work, tempo adjustment is employed, to change the speed of an input signal, without modifying the pitch value. As the proposed work involves different speakers who speak in a slow and fast manner, this step helps to maintain the required speed of a signal and also assist to reduce the speaking rate variation.
5) Silence Removal using Voice Activity Detection (VAD) Voice Activity Detection algorithm is widely used to identify the voiced and unvoiced region of a speech signal [18]. In general, a speech or speaker specific attributes are located in the voiced part, whereas the other undesirable components like silence or the background noise are located in the unvoiced part. Therefore, making useful discrimination between the voiced and unvoiced part can help to remove the irrelevant segment of a speech signal. The user/speaker usually takes a few seconds before and after saying a word while recording a speech utterance. So, the first and last 200 msec of a recorded speech signal might contain silence or irrelevant speech information. Therefore, silence in the beginning and at the end of speech file is removed to reduce the end point detection error.
The original input speech signal is passed through the above five steps. As a result, an effective pre-processed signal is obtained with better quality and intelligibility. Also, the proposed pre-processed signals are found to be louder and clear when compared with original signals. Figure 3 and Figure 4 show the spectrogram of the original and Pre-processed input speech signal for the word ―poojiam‖ uttered by four different speakers respectively. These Pre-processed signals are then used as an input for the modified feature extraction using GFCC technique.
B. Modified Feature Extraction using GFCC
Speech recognition performance can vary according to the type of feature extraction technique adopted for the particular application. The proposed methodology focuses on providing suitable features as an effective speech front-end using three different modified GFCC features. The subsequent sections explain the same in detail.
1) Feature Extraction using Multi Taper Yule Walker AR - GFCC (MTYW-GFCC)
An efficient feature extraction technique should extract the most useful spectral information from a signal, which can improve the recognition performance. This can be accomplished by using Power Spectral Density (PSD) estimation that extracts the frequency response of a signal. A power spectrum shows the amplitudes of all frequency components present in a segment of signal that can provide better discrimination between the speech segments. It helps to understand where the average power is distributed as a function of frequency.
Fig.3. Spectrogram of the Original Input Speech Signal for the Word
―Poojiam‖ Uttered by Four Different Speakers.
