This page contains LDM 2-phone classification result under matlab.

In this experiment, we use 1 dimensional Y[] and 5-dimensional X[].

Experiment steps:
1) choose a 50ms frame from raw speech files aa.snd and sh.snd, applying lpc-reconstruction to get speech-like synthetic signals aa and sh.
2) train ldm models for aa and sh by running EM recursion 500-3000 times. Therefore we get the ldm model parameters for aa and sh.
3) calculate likelihoods of signal aa and ldm model aa, signal aa and ldm model sh. Try to classify aa.
4) calculate likelihoods of signal sh and ldm model sh, signal sh and ldm model aa. Try to classify sh.

real speech signal and lpc-reconstructed signal [aa]

real speech signal and lpc-reconstructed signal [sh]

EM training 500 times [aa]

EM training 3000 times [sh]

Experiment result:
1) Likelihood value of signal aa and ldm model aa is 3053, likelihood value of signal aa and ldm model sh is -1552. Succeed to classify aa.
2) Likelihood value of signal sh and ldm model sh is 729, likelihood value of signal sh and ldm model aa is -320930. Succeed to classify sh.