Howard, Ian; (1991) Speech fundamental period estimation using pattern classification. Doctoral thesis (Ph.D), UCL (University College London).

Text Speech_fundamental_period_esti.pdf Download (15MB)

Abstract

The problem investigated concerns the robust estimation of fundamental period, only on the basis of representative speech pressure waveforms. The work has involved the design and development of a set of algorithms. The main intended application is in pattern processing acoustic and cochlear implant hearing aids. Essentially the task is to infer from the acoustic evidence available the points in time at which vocal fold closures occur. Its accomplishments both gives fundamental period information on a cycle-by-cycle basis and provides information concerning whether voicing is present. The task of detecting the point of closure of the vocal folds is formulated as a pattern recognition problem, and the pattern recognition technique employed uses the multi-layer perceptron (MLP). The first system configurations investigated were based on a preprocessing of the speech pressure waveform by a wide-band filterbank analyzer. This gave an input to the classifier which consisted of a set of adjacent time frames from the output of the filterbank. The output from the classifier was defined as being in one of two classes. In the first there is a period epoch marker at a given output frame, in the second there is not. This first classifier was trained to generate an output which signified the presence of a vocal fold closure at the centre of its input window. The fundamental periods between successive vocal-fold closures defined by these epoch markers, are given the name Tx. The labelling of both training and test data was performed semi-automatically by means of an algorithm that makes use of the output of a laryngograph. Developments of this first approach were then explored. These were primarily directed towards methods for reducing the training time for the MLP and improving the dme resolution of the fundamental period estimates. Different pre-processing stages were investigated and these included direct operation on the speech pressure waveform and the use of a simplified auditory filterbank. Methods to reduce the computation load required for practical implementation were examined and these resulted in a system using a low-order filterbank together with a smaller MLP network. The last configuration was of practical interest because it had a processing load small enough to be run in real-time on a portable DSP system. A real-time system was implemented in conjunction with Mr. John Walliker. First patient results using this system are reported following perceptual assessments made by Dr. Andrew Faulkner. A number of objective assessment techniques were developed and used to permit quantitative comparisons between fundamental period estimation algorithms to be carried out. These involved both quantitative comparisons between frequency contours and between time excitation epoch markers. Using these comparisons, various different configurations of the MLP-Tx algorithm were evaluated over a wide range of speakers and environmental conditions. The performance of the MLP-Tx algorithm was also compared against that of established fundamental frequency estimation algorithms, and its performance in competing noise was found to be better than that obtainable by the use of the peak-picking approach previously employed.

Download activity - last month

Export as JSONCSVXML

Download activity - last 12 months

Export as JSONXMLCSV

Downloads by country - last 12 months

Export as XMLCSVJSON

Archive Staff Only

View Item