In this thesis, we evaluate the front-end of Automatic Speech Recognition (ASR) systems, with respect to different types of spectral processing methods that are extensively used. A filter bank approach for front end spectral analysis is one of the common methods used for spectral analysis. In this work we describe and evaluate spectral analysis based on Mel and Gammatone filter banks. These filtering methods are derived from auditory models and are thought to have some advantages for automatic speech recognition work. Experimentally, however, we show that direct use of FFT spectral values is just as effective as using either Mel or Gammatone filter banks, provided that the features extracted from the FFT spectral values take into account a Mel or Mel-like frequency scale. It is also shown that trajectory features based on sliding block of spectral features, computed using either FFT or filter bank spectral analysis are considerably more effective, in terms of ASR accuracy, than are delta and delta-delta terms often used for ASR. Although there is no major performance disadvantage to using a filter bank, simplicity of analysis is a reason to eliminate this step in speech processing. These assertions hold for both clean and noisy speech.
展开▼
