Overview

Our speech research has concentrated on assuming that speech signals are time-stationary sequences. The advances based on this assumption have lowered the bound for speech research. This research now needs to focus on an area that basically defines speech as a time-varying signal, which can be analyzed in order to extract/add dimensions and features. Entropy within this extraction will allow us to more closely match the speech production system. There is a lower bound on entropy that can be achieved, and nonlinear methods will help us to accelerate the chances of reaching this lower bound.

The initial step in our process will be to demonstrate the capabilities of nonlinear modeling by implementing our basic nonlinear model, particle filtering, to our speech data. By implementing this model, we will be able to probe more attributes-extraction techniques, which will be used to recognize the speaker or implement the speaker-dependent speech recognition system. Our ultimate aim from being able to implement the speaker-dependent speech recognition system is to use far less features, as compared to the available systems, and to be able to have similar feature extraction techniques for speaker-independent speech recognition system.

Right now, our research is heading in the right direction, and we are receiving some positive results in random data. Yet, we still have to test our method on a speech database.

Our research falls under a fundamental research area and will not explore all the methods available for nonlinear modeling of the time-varying systems, but we will try to explore the potentials we have established by our nonlinear model in the other fields of research.

The benefit of such research is to find the newer areas of applied material for speech research and to fuel the exploration.

The plan that has been laid down:

  1. During the first year, we will explore more than one nonlinear method of speech databases for speaker verification. These results will be used to bench-mark our performance and then help us to redirect the research towards reduction in the features count.
  2. Our second year will be experienced through exploration of the details in reduction and will set a renewed bench-mark for speaker verification. We will utilize this bench-mark for speech verification.
  3. For our third year, we will fine tune our results and explore further possibilities of research in this field.