Fixed point arithmetic in speech recognition systems

Describes some of the methods used by the author and others in the construction of speech recognisers. The author has implemented some techniques using software written in a high level language, with special functions to implement the fixed point arithmetic processes. A number was represented as a 32 bit integer format, 16 bits for the whole part and 16 bits for the fraction. The techniques implemented were: Levinson-Durbin recursion for LPC calculation; Prime number fast Fourier transform; and cubic spline interpolation for linear time normalisation