I'm training a model for recognizing small set of commands in polish. I have created all nessesary files mentioned in tutorial, run training and got ~90% WER and SER accuracy using jfsg grammar. Before i have tried to use LM generated by LMP tool but got 100% error with it (not a signe word has been recognized). My audio data consists of 800+ audio files of commands spoken over and over by one person (about 0,8 hour).
I get only 2 sentences recognized, the rest of sentences is either falsely recognized as one of mentioned senteces or not recognized at all ("fsg_search.c", line 940: Final result does not match the grammar in frame 255).
My audio files are 16 bit little endian, 16kHz mono.I am using sphinxtrain, pocketshpinx and shpinxbase from github (however i have used latest sources downloaded from sourceforge )
The same data used with Kaldi ASR managed to train a model with a very small error (sub 10% in both SER and WER), so i'm suspecting something is wrong with my language model (both jfsg and lm), but i can't really pinpoint what it is.
What could be a possible cause of such a high error rate? Is my data/configuration garbage ?
Hello,
I'm training a model for recognizing small set of commands in polish. I have created all nessesary files mentioned in tutorial, run training and got ~90% WER and SER accuracy using jfsg grammar. Before i have tried to use LM generated by LMP tool but got 100% error with it (not a signe word has been recognized). My audio data consists of 800+ audio files of commands spoken over and over by one person (about 0,8 hour).
I get only 2 sentences recognized, the rest of sentences is either falsely recognized as one of mentioned senteces or not recognized at all ("fsg_search.c", line 940: Final result does not match the grammar in frame 255).
My audio files are 16 bit little endian, 16kHz mono.I am using sphinxtrain, pocketshpinx and shpinxbase from github (however i have used latest sources downloaded from sourceforge )
The same data used with Kaldi ASR managed to train a model with a very small error (sub 10% in both SER and WER), so i'm suspecting something is wrong with my language model (both jfsg and lm), but i can't really pinpoint what it is.
What could be a possible cause of such a high error rate? Is my data/configuration garbage ?
LMTool supports only English, you should have used srilm. Your current lm does not match the dictionary lowercase words.
Your jsgf grammar is not constructed correctly. I has has too many public rules, only the last one is actually used.
That was a silly question of me, sorry to waste your time on something that obvious,
Thank you kindly for your help.