Separation of Desired Speech from Interfering Speech Reverberating in a Room

A new algorithm has been developed to achieve the separation of desired speech from interfering speech reverberating in a room. The algorithm is developed by viewing this problem as a two-data channel adaptive noise cancellation problem. Using a least squares error criteria in the adaptive noise canceler approach leads to classical system identification problem. The solution estimates the room reverberant system and restores the desired speech by filtering the interfering speech with the estimated filter. The system identification problem is a well-known research field; however, this thesis project is characterized by two major features. One is the fact that it deals with a system involving a large number ( more than 1000 ) of parameters to be estimated. The other is that it handles non-stationary speech data signals, rather than just stationary noise signal. Of the many available techniques for system identification, a spectral analysis estimation method is best suited to this particular problem by virtue of its efficient computation, using the FFT, and the flexibility for applying it to any type of room. The proposed new algorithm is a spectral analysis estimation method, a modified technique for dealing with non-stationary speech signal. The Maximum Likelihood estimation technique directly derives this new algorithm by assuming a Gaussian colored noise for the desired speech, and by allowing the colored noise power spectral density to change frame by frame. The filtering process is realized by the overlap-save method. The algorithm uses finite length frames and is implemented in a recursive fashion, providing the capability of real-time processing and adaptability to possible changes of the room acoustic environment. In each frame, the filter estimate is updated on a frequency-by-frequency basis. The algorithm was implemented in a computer software program. Experiments using synthetic speech data were performed with two different types of room transfer functions, a 32 point delayed delta function, and a 1024 point room response experimentally measured in a room. The algorithm achieved above 20 DB signal-to-noise ratio in the output speech, when original speech signals with three different signal-tonoise ratio ( 6 DB, 0 DB, -12 DB ) were input into the algorithm. For actually recorded speech data, however, the result is too obscure to present in this thesis. More work needs to be done for this case. Thesis Supervisor: Bruce R. Musicus Title: Assistant Professor of Electrical Engineering & Computer Science