Overall Channels Weighting Factor

The input signal to the filter-and-sum module is typically a 16bit, 16KHz signal, and the output being treated by the diarization system is of the same characteristics. By using 16 bits it can represent values from -32767 to +32768 in a single channel in steps of 1 (resolution of the input). Such resolution gets modified when performing the weighted sum of $ N$ signals as the resolution becomes smaller than 1 (the range of possible values of the summed signal depends of the weights of the individual signals, it would be $ \frac{1}{N}$ for equal weighing). Although a higher resolution is available after the sum, the signal needs to be quantized to steps of unit value to fit it into the 16bit output channel, therefore getting a quantization error at each frame.

As the use of a signal output using more bits (like using floating points) creates an inconsistency with the standard signals used in the system and therefore was not considered as feasible, two simple modifications were done in order to minimize the amount of quantization error whenever possible. These are:

user 2008-12-08