If I have audio files recorded at 16Khz (mono) and want to decode tham with
pocketsphinx_batch with hub4wsj_sc_8k
acoustic model:
do I need to first sample-rate-convert these files to 8k before proceeding in order to get better accuracy ?
If yes, do I need to specify -samprate 8000 for running those sample rate converted files ?
Currently I'm running these files without giving any option for -samprate and am getting reasonably good accuracy
however when I run them with -samprate 8000, my accuracy is dropping
significantly
do I need to first sample-rate-convert these files to 8k before proceeding
in order to get better accuracy ?
No, pocketsphinx handles that. samprate means the rate of the audio, not the
rate for the model. Since your rate is already 16k, you dont' need anything
If yes, do I need to specify -samprate 8000 for running those sample rate
converted files ?
No, see above
Currently I'm running these files without giving any option for -samprate
and am getting reasonably good accuracy
Just FYI: I tried downcoverting to 8k and decoding between the time I posted
the query and got your reply. Results are in
synch with what you have mentioned (No difference between "results of original
16khz file decoded without any
-samprate optiion" and " 8k downconverted files with -samprate 8000 option". Downconveted file when decoded
without -samprate options are giving very poor recognition accuracy".
Regards,
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Hi,
If I have audio files recorded at 16Khz (mono) and want to decode tham with
pocketsphinx_batch with hub4wsj_sc_8k
acoustic model:
however when I run them with -samprate 8000, my accuracy is dropping
significantly
Note: I got this doubt after looking at the post https://sourceforge.net/proj
ects/cmusphinx/forums/forum/5471/topic/3744474 by
geetg and thought of reconfirming this.
Regards,
No, pocketsphinx handles that. samprate means the rate of the audio, not the
rate for the model. Since your rate is already 16k, you dont' need anything
No, see above
That's great
In that post audio files were sampled at 8k
Thanks nshmyrev.
Just FYI: I tried downcoverting to 8k and decoding between the time I posted
the query and got your reply. Results are in
synch with what you have mentioned (No difference between "results of original
16khz file decoded without any
-samprate optiion" and " 8k downconverted files with -samprate 8000 option". Downconveted file when decoded
without -samprate options are giving very poor recognition accuracy".
Regards,