International Scientific Journals
of Scientific Technical Union of Mechanical Engineering "Industry 4.0"

In this paper we will discuss an important topic such as the WER (word error rate). Imagine if we use an ASR system in a real event for approximately one hour or more. We would have a lot of issues like: the quality of the transcription of words, the time of the processing of the words spoken. And so a lot of WER will be produced by this event. The high rate of ASR errors have demanded the necessity to find better techniques in order to correct such errors. The improvement of the system it is a necessity not only to have a better stability of the system but also to prevent the high costs in order to use our resources in the best way as possible.

In this paper we will use different feature extraction approach in order to compare the stability of the Automatic Speech Recognition (ASR) on different SNR noise, in different situations. By using the classic feature extraction like: Mel filter Bank, Fast Fourier Transformation (FFT), Discrete Fourier Transformation (DFT). After we implement this methods for the different situations then we will warp the frequency of this methods directly on the spectrum in order to see if we get better results. The results will be compared with one another using the Word Error Rate algorithm (WER).

In this paper we evaluate the growth of Automatic Speech Recognition systems in respect to the various forms of spectral analysis ways used. A straightforward analysis of platter and Gammatone filter banks used for spectral analysis compared with the direct use of FFT spectral values is taken into account. This analysis was supported understanding the effectiveness of existing Automatic Speech Recognition systems that are specifically targeted on platter and Gammatone filter banks compared with FFT spectral values. We discover that warping the FFT spectrum directly, instead of using filter bank averaging, provides an additional precise approximation to the sensory activity scales. Direct use of FFT spectral values are even as effective as using either Gammatone or Linear Prediction filter banks, as long as the feature extracted from the FFT spectral values takes into consideration a Gammatone or platter like frequency scale. Computing speech signals using FFT or filter bank spectral features and utilizing a method supported by a sliding block of spectral features, is shown to be simpler in terms of ASR accuracy.