Speech Inverse Filtering by Simulated Annealing Algorithm

SPEECH INVERSE FILTERING BY SIMULATED ANNEALING ALGORITHM Chang-Shiann Wu Department of Information Management Shih Chien University Kaohsiung, Taiwan, R.O.C. Email: chwu4142@ms5.hinet.net Y u-Fu Hsieh Chung Shan Institute of Science and Technology Lung Tan Tao Yuan, Taiwan, R.O.C. Email: aliceufo@msl.hinet.net ABSTRACT The purpose of this study is to develop one solution to the speech inverse filtering problem. A new efficient articulatory speech analysis scheme, identifying the articulatory parameters from the acoustic speech waveforms, was induced. The algorithm is known as simulated annealing, which is constrained to avoid non-unique solutions and local minima problems. The constraints are determined by the articulatory-to-acoustic transformation function and the boundary conditions for the articulatory parameters. The cost function is defined as a percentage of the weighted least-absolute-value error distance between the first four formant frequencies of the articulatory model and the first four formant frequencies determined from speech analysis. It is used to optimize the vocal tract parameters to match a specified set of formant characteristics. A 1% error criterion was found to be both practical and achievable. 1. INTRODUCTION Articulatory synthesis is the production of speech sounds using a model of the vocal tract, which directly or indirectly simulates the movements of the speech articulators. It provides a means for gaining an understanding of speech production and for studying phonetics. Articulatory synthesis usually consists of two separate components. In the articulatory model, the vocal tract is divided into many small sections and the corresponding cross-sectional areas are used as parameters to represent the vocal tract characteristics. In the acoustic model, each cross- sectional area is approximated by an electrical analog transmission line to simulate the speech sound propagation through the vocal system as well as the physics of the physiological-to- acoustic transformation. To simulate the movement of the vocal tract, the area functions must change with time. Each sound is designated in terms of a target configuration and the movement of the vocal tract is specified by a separate fast or slow motion of the articulators. The recovery of articulatory movements from the speech signal, known as the speech inverse filtering problem, is difficult due to the non-uniqueness of the solution. This problem has been the subject of research for several applications, including articulatory synthesis, speech recognition, low-bit-rate speech coding, and text-to-speech synthesis. Here we attempt a new solution using the simulated annealing algorithm, which is a "constrained multidimensional nonlinear optimization problem." The coordinates of the jaw, tongue body, tongue tip, lips, velum, and hyoid compose the multidimensional articulatory vector. A comparison between the model-derived and the target- frame first four formant frequencies forms the cost function. There are two constraints: (1) the articulatory-to-acoustic transformation function, and (2) the boundary conditions for the articulatory parameters. The optimum articulatory vector is obtained by finding the minimum cost function. Once the optimum articulatory vector is determined, the articulatory model -231- determines the vocal tract cross-sectional area function which in turn is used by the articulatory speech synthesizer. 2. IMPLEMENTATION OF THE ARTICULATORY MODEL Geometric data concerning the …

Content truncated to 500 words