The effects of reduced precision bit lengths on feedforward neural networks for speech recognition

An investigation of using fixed-point arithmetic on a neural networks is presented. A formula that estimates the standard deviation of the output differences of fixed-point and floating-point networks is developed. The formula provides a priori knowledge regarding the required number of bit precision that should be employed to achieve acceptable recognition rate on the feedforward recall phase. A time delay neural network (TDNN) with speaker independence is used to do unvoiced stop consonants recognition, namely P, T, K. The fixed-point arithmetic implementation offers comparable simulation results to that of a floating-point implementation. The recognition rate for the test set employing fixed-point arithmetic in both training and recall is between 75% and 90%. A single digit speaker independent problem is also investigated to prove the validity of formula further