| Speech recognition is one of the most promising technologies of the future as seen from its many contributions in applications such as mobile communications, medical devices, wearable computing, and general home appliances. Speech recognition systems provide a wide range of different uses including voice-enabled web pages to enhance the virtual world experience, voice-activated control systems in cars for Internet access, emergency assistance, or navigation, and speech-to-text systems that can translate voice commands into document text. The incorporation of speech recognition technology into our everyday lives has prompted continuous research efforts to develop a system that is capable of not only recognizing speech, but also understanding and responding accurately to speech. Ideally, a speech recognition system should be capable of performing spoken language processing which is the ability to accurately recognize, understand, and ultimately, respond to speech inputs. Current systems lack the ability to perform spoken language processing because they are unable to completely represent features in speech due to speech and speaker variability. Systems must implement feature extraction techniques that can accurately represent speech in order to preserve variability and provide high performance speech systems.; In this thesis, a wavelet-based approach to speech recognition is presented. Wavelet analysis is a mathematical tool that provides a multiple scale approach to speech signal representations. Moreover, optimal wavelet methods can be designed to capture the naturalness of speech, as well as, provide feature extraction for speech recognition systems. Computer simulations and hardware analyses have shown that wavelet feature extraction not only performs well relative to conventional methods, but it also provides a mathematical framework for implementing complex speech systems with greater design flexibility.; The major contributions of this thesis include the design and implementation of optimal wavelet methods for feature extraction, and its analysis in a complete speech recognition system. In addition for real-time speech recognition systems, algorithms most often have to be implemented in hardware. Hence, hardware evaluations of wavelets for speech recognition systems using reconfigurable hardware architectures are also explored. |
