Using Nbest crashes Windows Phone example

[Toine db]

I'm trying to implement NBest output as an extension for the Windows example but App keeps crashing when using "ps_nbest", "ps_nbest_next" and "ps_nbest_hyp". But the App keeps crashing when these methods are used!

I didn't use or create any special things, its almost an exact copy of the write_nbest() from PocketSphinx_Batch.
I'm testing it on a WindowsPhone with the digits.gram, that works fine when only using ps_get_hyp.

Are there any known issues with ps_nbest and pocketsphinx in Windows platform???

the code I added for nBest, with this the Apps crashes

ps_nbest_t *nbest;
        int32 i, n, score;
        const char* nbestHypothesis;

        auto nbestResults = ref new Vector<Platform::String^>();
        n = 5;

        nbest = ps_nbest(decoder);
        for (i = 0; i < n && nbest && (nbest = ps_nbest_next(nbest)); i++)
        {
            nbestHypothesis = ps_nbest_hyp(nbest, &score);
            auto nbestHyp = convertCharsToString(nbestHypothesis);
            nbestResults->Append(nbestHyp);
        }

        if (nbest)
        {
            ps_nbest_free(nbest);
        }

        OnNbestResultFinalizedBySilence(nbestResults);

PS: When (quickly) debugging I seen that it passes through "ps_nbest" the first time, but with a very long dela(10 seconds) and the second time it uses "ps_nbest" the app crashes. I'll try to get more logs when my system is up and running again.

[Toine db]

I have done some debugging and the first thing I notice is that nbest = ps_nbest(ps); takes more then 5 seconds, when decoded a simple digit.

How can nBest be so slow on a Phone, while PocketSphinx_Batch does this within miliseconds?

FYI: I solved/found the crash source, turns out that ps_nbest_hyp() sometimes returns a NULL that I didn't expect.

[Nickolay V. Shmyrev]

Nbest speed depends on lattice size. If lattice is huge it may take quite some time. Lattice size is printed in the log and controlled by beam parameters.

[Toine db]

OK, thanks for the reply.

I'll try something else as well; if the nbest is faster when using complete audio bytes for a single utterance (like ps_batch), instead of continously reporting audio bytes like currently is done in the example (like ps_continous).

[Toine db]

Again thanks for you support.

I added NBest example to the Windows Phone project.
https://github.com/cmusphinx/pocketsphinx-wp-demo/pull/5

It includes a lot readbility improvements and cleanup and memory leak fixes etc.

PS: working on antother Pull request to improve the Struct used by the interface.
(instead of string thats needs to be split)