Adaptive algorithms and structures with potential application in reverberation time estimation in occupied rooms

Realistic and accurate room reverberation time (RT) extraction is very important in room acoustics. Occupied room RT extraction is even more attractive but it is technically challenging, since the presence of the audience changes the room acoustics. Recently, some methods have been proposed to solve the occupied room RT extraction problem by utilizing passively received speech signals, such as the maximum likelihood estimation (MLE) technique and the artificial neural network (ANN) scheme. Although reasonable RT estimates can be extracted by these methods, noise may affect their accuracy, especially for occupied rooms, where noise is inevitable due to the presence of the audience. To improve the accuracy of the RT estimates from high noise occupied rooms, adaptive techniques are utilized in this thesis as a preprocess ing stage for RT estimation. As a demonstration, this preprocessing together with the MLE method will be applied to extract the RT of a room in which there is significant noise from passively received speech signals. This preprocessing can also be potentially used to aid in the extraction of other acoustic parameters, such as the early decay time (EDT) and speech transmission index (STI). The motivation of the proposed approach is to utilize adaptive techniques, namely blind source separation (BSS) and adaptive noise cancellation (ANC), based upon the least mean square (LMS) algorithm, to reduce the noise level contained in the received speech signal, so that the RT extracted from the signal output generated by the preprocessing can be more accurate. Further research is also performed on some fundamental topics re lated to adaptive techniques. The first topic is variable step size LMS (VSSLMS) algorithms, which are designed to enhance the convergence rate of the LMS algorithm. The concept of gradient based VSSLMS algorithms is described, and new gradient based VSSLMS algorithms are proposed for applications where the input signal is statistically stationary and the signal-to-noise ratio (SNR) is zero decibels or less. The second topic is variable tap-length LMS (VTLMS) algorithms. VTLMS algorithms are designed for applications where the tap-length of the adaptive filter coefficient vector is unknown. The target of these algorithms is to establish a good steady-state tap-length for the LMS algorithm. A steady-state performance analysis for a VTLMS algorithm, the fractional tap-length (FT) algorithm is therefore provided. To improve the performance of the FT algorithm in high noise conditions, a convex combination approach for the FT algorithm is proposed. Furthermore, a new practical VTLMS algorithm is also designed for applications in which the optimal filter has an exponential decay impulse response, commonplace in enclosed acoustic environments. These original research outputs provide deep understanding of the VTLMS algorithms. Finally, the idea of variable tap-length is introduced for the first time into the BSS algorithm. Similar to the FT algorithm, the tap-length of the natural gradient (NG) algorithm, which is one of the most important sequential BSS algorithms is also made variable rather than fixed. A new variable tap-length NG algorithm is proposed to search for a steady-state adaptive filter vector tap-length, and thereby provide a good compromise between steady-state performance and computational complexity. The research recorded in this thesis gives a first step in introducing adaptive techniques into acoustic parameter extraction. Limited by the performance of such adaptive techniques, only simulated studies and comparisons are performed to evaluate the proposed new approach. With further development of the associated adaptive techniques, practical applications of the proposed approach may be obtained in the future