Modifying pocketsphinx to do livepretend

[Mike Medved]

Hi...

I'm looking to take pocketsphinx still in batch mode (reading from a file), but instead of processing it all at once, I want to process it some chunk of data at time. I made a function to replace ps_decode_raw that streams data in. It looks like this:

static int medvedsDecode(ps_decoder_t ps, FILE rawfh,
char const *uttid, long maxsamps)
{
long total, pos, i;

ps_start_utt(ps, uttid);

/* Otherwise decode it in a stream. */
total = 0;
while (!feof(rawfh)) 
{
    long long start = 0, end = 0;
    long difference = 0;

    int16 data[256];  
    size_t nread;

    //ClockTime(CLOCK_REALTIME, NULL, &start);

    nread = fread(data, sizeof(*data), sizeof(data)/sizeof(*data), rawfh);
    ps_process_raw(ps, data, nread, FALSE, FALSE);
    total += nread;

    //ClockTime(CLOCK_REALTIME, NULL, &end);

    //difference = (end - start) / 1000;  //(in microsecs)

    //usleep((10000 * (512 / 80)) - difference);         
}

ps_end_utt(ps);

return total;

}

I'm not really worried about the sleeping portion yet - what is weird (and what I'm looking for help on) is that the first file the decoder processes has really bad accuracy (say 50% WER). Then everything after that works well. If I run the same file twice in a row, the first one will be bad, but the second time it is fine (0% WER). Here is an example from the align file:

*** close *** help *** eva *** go to sleep *** home page *** I need help *** next step page back page forward previous step wakeup (ALL-FINLEYH\AUDIO\ALL\TRACK-1)
GO close BACK help GO eva TO go to sleep GO home page BACK GO need help GO next step page back page forward previous step wakeup (ALL-TRACK-1)
Words: 20 Correct: 19 Errors: 8 Percent correct = 95.00% Error = 40.00% Accuracy = 60.00%
Insertions: 7 Deletions: 0 Substitutions: 1
close help eva go to sleep home page i need help next step page back page forward previous step wakeup (ALL-FINLEYH\AUDIO\ALL\TRACK-1)
close help eva go to sleep home page i need help next step page back page forward previous step wakeup (ALL-TRACK-1)
Words: 20 Correct: 20 Errors: 0 Percent correct = 100.00% Error = 0.00% Accuracy = 100.00%

Thoughts? Pointers? I've tried different numbers for the data size (256 in the code above), with varying results, but always that first file is pretty awful.

[Mike Medved]

Also, I'm curious why it seems that full batch mode (i.e. process the whole file at once) seems to have significantly better accuracy than a streaming batch mode. Do the algorithms differ?

[Nickolay V. Shmyrev]

If you are using wsj1, you need to set cminit to something around 50, default value is bad for the models with -transform dct.