Speech processing and recognition using artificial intelligence methods

Many problems related to analysis and recognition of sound signals are characterized by the fact that finding the proper analysis rule or the proper recognition algorithm is very difficult. In addition the signal analysis methods and the methods of its recognition must be strictly adapted to the specific features of the particular task considered. It is worth noticing, that the situation is essentially different from the one encountered in the sound processing. The method of lowand high-pass filtering, compression techniques or algorithms of spectral transformations for a sound signal do not actually depend on the type of signal or the purpose for which it is being registered and processed. Therefore in the field of sound signal processing an enormous progress has been achieved, and the elaborated methods are to a high degree universal. The effect is additionally enhanced by the availability of affordable [lowcost] and convenient DSP technique. On the contrary in the tasks of sound signals analysis and processing the progress is much slower and the unification of methods and standardization of algorithms encounters considerable difficulties. The main source of these difficulties is the fact that in almost every task of signal analysis the features to be extracted and the required signal parameters are different, they are strongly dependent on the specific task being solved and they are expected to provide answers for different questions. Similarly in the tasks of sound signal recognition the criteria and goals of their classification can be very different even for the same signal types. Yet, the approach unification in the above mentioned field is strongly recommended, because it enables[promotes] more cost-effective and faster development of the required solutions for particular problems. It seems that in the field of analysis and recognition of sound signals a very promising direction in the search for such unification, and the universal solutions which might lead to it, are the methods used in the artificial intelligence research. There are several definitions of artificial intelligence [1,2], but in the tasks of speech processing and recognition the most appropriate seems to be the slightly narrower concept of computational intelligence [1]. In the discussed field of interest (speech as a biomedical signal) the methods of artificial intelligence are mostly employed to: preliminary signal processing and filtering [6], determination of the space of phono-acoustic features and its visualization [5], speech recognition, identification of the speaker or pathological states [3,4,7], understanding the signal [8]. In addition to classical methods of pattern recognition, fuzzy systems or genetic algorithms the neural networks seem to deserve a special attention. The neural network MAVEBA 2001, Firenze, Italy 228 ISCA Archive http://www.isca-speech.org/archive Models and Analysis of Vocal Emissions for Biomedical Applications (MAVEBA) 2 nd International Workshop