testing simon

A. I am testing sam. I am not sure about the jconf file (error message: “Couldn't open julius jconf file: "".“. Where is it located? When I am searching my computer, I find adin.jconf, default.jconf, Sample.jconf (each configuration file is located at a different location). Which one should I choose?

1. I tested sam with Sample.jconf, and it seemed to work. But to be honest: Isn’t this just a dummy file? Every line begins with an # (number sign).

2. adin.jconf is very short, and everything is commented out.

3. default.jconf is probably the correct choice because some lines are valid, e.g.:

[...]
-h hmmdefs
[...]
-hlist tiedlist
[...]
-penalty1 5.0 # first pass
-penalty2 20.0 # second pass
[...]

Location of this file: /usr/share/kde4/apps/simond/default.jconf
I think that I will use sam with this file.

B. I assume that sam > Build model is the same like simon > Synchronize because the button is the same (green circular arrow).

This means that I can use simon for recording new words, and synchronize the speech model. In my last video, all words (more than 200 words) were recognized correctly. How can I eliminate wav files that are not so good? I need an efficient way to fix future problems. sam seems to fill the gap. With simon, I can record words, and synchronize. With sam, I can find out which words are below 100% confidence score. These words can be edited with sam (very nice feature).

So, simon is good for the main work. Fixing bad wav files can be done with sam. E.g. I found a wav file with a confidence score of about 89%. I edited this wav file with sam (sam offers the option to re-record a wav file). I will see whether this improves the confidence score after rebuilding (sam > Build model), and testing the model (sam > Test model).

C. It seems that sam is working without ksimond.

D. I don’t know yet how to handle wav files with multiple words. With sam, the confidence score of these wav files is 0 %. I added with simon the grammar structure “Unknown Unknown Unknown Unknown“, but it still doesn’t work out. But now, I found one sample which had a confidence score of 50%:

Result 6 of 10
=====================
Sentence: verlangsamende verlangsamendem verlangsamenden verlangsamendes
SAMPA:
Raw SAMPA:
Average Confidence: 50
Confidence Scores: 1.72309e-12 100 4.8642e-05 100

So I can say: it seems to work when I am adding the sentence structure “Unknown Unknown Unknown Unknown“.

E. The word deutschfeindlichen has a confidence score of 90.0133 %:

Result 1 of 3
=====================
Sentence: deutschfeindlichen
SAMPA:
Raw SAMPA:
Average Confidence: 90.0133
Confidence Scores: 90.0133

Result 2 of 3
=====================
Sentence: deutschfeindlichem
SAMPA:
Raw SAMPA:
Average Confidence: 9.98674
Confidence Scores: 9.98674

Result 3 of 3
=====================
Sentence: deutschfeindliche
SAMPA:
Raw SAMPA:
Average Confidence: 8.29019e-17
Confidence Scores: 8.29019e-17

So the alternatives deutschfeindlichem and deutschfeindliche have a much lower confidence score. This is fine. I can see how good it works internally. Because in the video (see link above), you can see 100 % perfect results. But internally, it is just about 90 % for deutschfeindlichen.

Here is another example Wortbreite:

Result 1 of 2
=====================
Sentence: Wortbreite
SAMPA:
Raw SAMPA:
Average Confidence: 0.0408209
Confidence Scores: 0.0408209

Result 2 of 2
=====================
Sentence: Wortbreiten
SAMPA:
Raw SAMPA:
Average Confidence: 99.9592
Confidence Scores: 99.9592

It should have recognized Wortbreite, but it recognized with more than 99 % confidence score the word Wortbreiten. So this match is wrong.

F. Conclusion: I hope you got an impression of sam. sam seems to be a great tool for speech model development.

Tags: jconf, sam

This entry was posted on Tuesday, December 29th, 2009 at 6:26 am and is filed under Ubuntu. You can follow any responses to this entry through the RSS 2.0 feed. Both comments and pings are currently closed.