AM scores per frame per senone

[kinoko]

Is there a program/script that outputs AM scores for each frame and each senone for input audio?
I inherited an acoustic model, trained with sphinxtrain.
I am okay with using/modifying C if that is needed, I just not sure where to start.

gbr

[Nickolay V. Shmyrev]

Relevant pocketsphinx options are

-compallsen     no      Compute all senone scores in every frame (can be faster when there are many senones)
-senlogdir              Directory to log senone score files to

You can also check acmod.c in pocketsphinx

[kinoko]

Thanks!

It works with 5prealpha.
But I got Failed to calibrate voice activity detection for v0.8. I try to turn off VAD (-remove_silence no), but according to source, cont_ad_calib is called anyhow.

As for the dump file, data stream starts after the header and after BYTE_ORDER_MAGIC 44 33 22 11. Somehow filesize does not add up in longer files, but I have to check again.
The number of scores is not fixed actually... (yey). For those who are interested, the relevant part:

int
acmod_write_scores(acmod_t *acmod, int n_active, uint8 const *active,  int16 const *senscr, FILE *senfh)
/* Uncompressed frame format:
     *
     * (2 bytes) n_active: Number of active senones
     * If all senones active:
     * (n_active * 2 bytes) scores of active senones
     *
     * Otherwise:
     * (2 bytes) n_active: Number of active senones
     * (n_active bytes) deltas to active senones
     * (n_active * 2 bytes) scores of active senones
     */