Hello,
I made a new acoustic and language model for the Greek language. The recognition accuracy is really good but I have one problem. It takes about 1 minute for the decoder to load the whole model with pocketsphinx, while both the acoustic and the language models are not large (acoustic model folder: 4.1MB, lang model .lm: 23.5KB and lang model .lm.bin: 793.5KB).
The decoder loads very fast some models I downloaded (such as the default en-us model). Also the sphinx4 decoder loads my model very fast (strange, uh?).
Also, in the initialization of the decoder I get the following warning: WARNING: "hash_table.c", line 150: Very large hash table requested (1095019 entries).
I searched on the Internet extensively and found nothing about my problem. Could you tell me what's going on?
Last edit: Eleni Adamantidou 2018-08-19
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Thank you!!!
The dictionary was too large. Now everything's ok.
But could you explain me something?
I initially used the dictionary I found in the model for the Greek language I downloaded. How can pocketsphinx load this model quickly without problems, while it loads this dictionary combined with my language and acoustic models so slowly? It's like it loads the same dictionary slower in the second case.
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Hello,
I made a new acoustic and language model for the Greek language. The recognition accuracy is really good but I have one problem. It takes about 1 minute for the decoder to load the whole model with pocketsphinx, while both the acoustic and the language models are not large (acoustic model folder: 4.1MB, lang model .lm: 23.5KB and lang model .lm.bin: 793.5KB).
The decoder loads very fast some models I downloaded (such as the default en-us model). Also the sphinx4 decoder loads my model very fast (strange, uh?).
Also, in the initialization of the decoder I get the following warning: WARNING: "hash_table.c", line 150: Very large hash table requested (1095019 entries).
I searched on the Internet extensively and found nothing about my problem. Could you tell me what's going on?
Last edit: Eleni Adamantidou 2018-08-19
Probably your dictionary is too large, you'd better reduce it to smaller one restricting to the words actually used in langauge model
Other items to check to speedup loading:
a) You are using PTM model instead of continuous one
b) You converted mdef file to binary format with pocketsphinx_mdef_convert
Thank you!!!
The dictionary was too large. Now everything's ok.
But could you explain me something?
I initially used the dictionary I found in the model for the Greek language I downloaded. How can pocketsphinx load this model quickly without problems, while it loads this dictionary combined with my language and acoustic models so slowly? It's like it loads the same dictionary slower in the second case.
I think initially it was also slow, if you are still curious you can try to reproduce the problem and run under oprofile to see where it spends time.