The matched filter is the implementation used for this
speaker verification system. The initial consideration for an
implementation was to simply matched filter the input waveforms of two
signals as opposed to certain characteristics of these waveforms. This
implementation consisted of taking two input signals and cutting out
the sections of the waveforms where speech existed (based on a visual
determination). Then, one waveform was scaled to the length of the other,
the two signals were normalized, and the dot product was computed. This
implementation presents various, obvious problems. The most significant
problem encountered is aligning the phase of the two input signals. This,
it turns out, is a very difficult obstacle to overcome. As a result, this
implementation was deemed impractical very quickly. The consideration of
several other problems experienced in a speaker verification system led to the
development of the final implementation.
LOSING - "If at first you don't succeed, failure may be your style."