I would like to use Sphinx4 with a little french dictionnary (no more than 10 words for the moment) on an iMac with Mac OSX 10.4.11.
First, following the Robust Group Tutorial, I ran the training/decode process with my own .dic, .filler, .phone etc... files and with my own audio files in .raw little endian format (as in big endian format, it did not work) at 16000 Hz.
After some tries, all ran successfully and I even got:
SENTENCE ERROR: 0.0% (0/14) WORD ERROR RATE: 0.0% (0/70)
Then I enclosed my new models in a .jar file, in every way similar to TIDIGITS_8gau_13dCep_16k_40mel_130Hz_6800Hz and I used it with an adapted version of HelloDigits. Once more, all ran successfully but each of my words was decoded as a <sil>.
To verify I ran the original version of HelloDigits with TIDIGITS_8gau_13dCep_16k_40mel_130Hz_6800Hz and this time each word was not decoded as a <sil> but misunderstood.
Could anyone tell me why it does not work and what I can do to overcome this problem?
Thanks in advance.
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Well, the issue is big-endian for sure, it's like shooting yourself in a leg. But to overcome this problem I suggest you to try to check everything. You need to add a DataDumper in a frontend pipeline and try to compare the numbers. Also you need to get score output. By comparision of results of little-endian and big-endian machine you could find the issue.
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Hi,
I would like to use Sphinx4 with a little french dictionnary (no more than 10 words for the moment) on an iMac with Mac OSX 10.4.11.
First, following the Robust Group Tutorial, I ran the training/decode process with my own .dic, .filler, .phone etc... files and with my own audio files in .raw little endian format (as in big endian format, it did not work) at 16000 Hz.
After some tries, all ran successfully and I even got:
SENTENCE ERROR: 0.0% (0/14) WORD ERROR RATE: 0.0% (0/70)
Then I enclosed my new models in a .jar file, in every way similar to TIDIGITS_8gau_13dCep_16k_40mel_130Hz_6800Hz and I used it with an adapted version of HelloDigits. Once more, all ran successfully but each of my words was decoded as a <sil>.
To verify I ran the original version of HelloDigits with TIDIGITS_8gau_13dCep_16k_40mel_130Hz_6800Hz and this time each word was not decoded as a <sil> but misunderstood.
Could anyone tell me why it does not work and what I can do to overcome this problem?
Thanks in advance.
Well, the issue is big-endian for sure, it's like shooting yourself in a leg. But to overcome this problem I suggest you to try to check everything. You need to add a DataDumper in a frontend pipeline and try to compare the numbers. Also you need to get score output. By comparision of results of little-endian and big-endian machine you could find the issue.