pocketsphinx recognition not consistent

[Jagadeesh Balam]

When I loop the ps_decoder to recognize the same sentence , I get different output each time. I am doing the following

for(i=1;i<argc;i++){

    fh = fopen(argv[i], "rb");
nsamp = fread(buffer, 2, 80000, fh);
    rv = ps_start_utt(ps, itoa(rand(),uttname,10));

   rv = ps_process_raw(ps, buf, length, FALSE, FALSE);
   hyp = ps_get_hyp(ditcKwsBeObject->ps, &score, &uttid);

}

When I give the same filename multiple times as the input argument, I am not getting the same hypothesis all the time. Is this expected, or am I not clearing some memory? Shouldn't ps_start_utt take care of clearing all the memory corresponding to the previous utterance.

Thanks in advance for any insights.
Jagadeesh

[Nickolay V. Shmyrev]

Yes, this discussed many times that dither uses random values, you can disable it if it's a problem for you.

[David Huggins-Daines]

Hi - it is more likely that this might be the effect of cepstral mean normalization when running in live mode. It uses a weighted average of past frames which gets updated after each utterance.