This Small Business Technology Transfer (STTR) Phase I project develops a model for recognizing spoken words based upon principles of neural computation. By exploiting the presumed role of nervous-system rhythms in neural computation, a model of time-frequency integration of signals that are a few hundreds of milliseconds long (e.g. whole words) will be developed. The project targets the evaluation of a model for recognizing diphones -- speech segments of duration of few tens of milliseconds. This model utilizes a template matching circuit (TMC) inspired by presumed principles of cortical neural processing, with a sub-threshold gamma oscillatory input with a frequency of about 30 Hz at its core. One property of the TMC is insensitivity to time-scale variations of the input stimuli. Such a property is needed to recognize speech tokens that inherently exhibit phonemic variability. The specific aims of Phase I are: (1) to quantify the TMC performance as a function of interacting rhythm frequencies and neuronal parameters such as time constants or threshold of firing, and (2) to compare the TMC performance to that of other methods of template matching, such as dynamic time warping (DTW) and static neural networks.
Speech recognition systems have evolved to reach reasonably performance in recent years, however the implementation techniques are clearly very far removed from the approach taken by the human brain. There are multiple benefits in attempting do develop biologically inspired models, for example to produce potentially more accurate and robust word recognition.
