testing simon

How I build Ralf's German dictionary version 0.2.1: (more…)

How I create Ralf's German dictionary version 0.2:

1. Add this rule to improve-german.xsl:

<xsl:when test="contains(lower-case(../grapheme), 'prüf')"><xsl:value-of select="replace($sierra, 'pʀyf','pʀyːf')"/></xsl:when>

2. Add rule: (more…)

Take a look into Ralf's German dictionary:

<lexeme role="Substantiv">
<grapheme>Packpapieren</grapheme>
<phoneme>pakpapiːʀən</phoneme>
<phoneme>pakpapiːɐ̯n</phoneme>
</lexeme>

Take a look into the corresponding section of ralfdic-scenario-20100725.xml:

<word>
<name>Packpapieren</name>
<pronunciation>p a kp a p i: R @ n</pronunciation>
<terminal>Substantiv</terminal>
</word>
<word>
<name>Packpapieren</name>
<pronunciation>p a kp a p i: ah n</pronunciation>
<terminal>Substantiv</terminal>
</word>

What is wrong? The phoneme kp should be written as two separate phonemes.

This is the current situation: I increased the size of the active vocabulary from 18000 to 300000. When I export my current scenario, I get an XML file of 58 MB with the name ralfdic-scenario-20100725.xml (please compare this value with the size of my phonetic dictionary: Ralf's German dictionary 0.1.9.9 has a size of 45 MB). This means that almost all words that are included in Ralf's German dictionary are included in my current simon scenario. Only the words that are affected are excluded from the active vocabulary (just about 1000 words are affected. This means that almost all words that you can find in Ralf's German dictionary can be found in ralfdic-scenario-20100725.xml (license: GPLv3).

But what happens when I dictate? simon obviously recognizes only a subset of 18000 words (only words that are included in Ralf's German IPA FLAC files). Why is that?

Take a look into file:///home/ubuntu/.kde/share/apps/simon/model/prompts (license of this file is GPLv3). My current prompts file contains only 18000 words, and not 300000 words. But my German scenario file ralfdic-scenario-20100725.xml contains 300000 words. And when I dictate, simon only recognizes words that are included in the prompts file. simon doesn’t recognize the other words that are part of the active vocabulary.

My current prompts file is a subset of my current active vocabulary.
My current active vocabulary is a superset of my current prompts file.

How can I solve this issue? I want that simon recognizes words that haven’t been trained before. Even words that are marked in red color in the active vocabulary should be recognized. I don’t want to record all 300000 words. This issue should be solvable, but how?

At the moment, Ralf’s German speech model 0.1.3 is installed on my computer (as user generated model). I would like to increase the vocabulary size (from 18000 words to 300000 words). So I am trying the following: I import Ralf's German dictionary 0.1.9.9. This will create of course a lot of duplicate entries in the active vocabulary.

This is what I am doing after the import of the dictionary:
- add terminals to the grammar (Adjektiv, Substantiv, Verb),
- press the Synchronize button.

So this is the situation at the moment:
Ralf's German speech model 0.1.3 contains 18000 words (it has been trained with 18000 audio files). I want to see how many additional words I have to train to cover all words that are contained in Ralf's German dictionary.

Then the following error message appears:

I could use a hint what I should do to increase the vocabulary size without having to train all the words. Do you have any suggestions?

Edit: I try the following: I clear the whole active vocabulary. Then I import Ralf's German dictionary 0.1.9.9 again (as active vocabulary). Then I disconnect simon. Then I restart ksimond. Then I press the simon connect button. Then I press the Synchronize button.

Now I have to wait a few minutes or so. The computer reacts pretty slowly because obviously simon is using a lot of processing capacity.

It was possible to compile the speech model. Now I press the Activate button. Then the following error message appears:

The recognition reported the following error:
The recognition could not be started because your model contains words that consits of sounds that are not covered by your acoustic model.

You need to either remove those words, transcribe them differently or train them.

Warning: The latter will not work if you are using static base models!

This could also be a sign of a base model that uses a different phoneme set than your scenario vocabulary.

The following words are affected (list may not be complete):
abelschen, abelschen, abonnierbaren, abonnierbaren, abonnierende, abonnierende, abonnierendem, abonnierendem, abonnierender, abonnierender, abonnierendes, abonnierendes, adressatengerechten, adressate…

The following phonemes are affected (list may not be complete):
*-C+e or biphone C+e, *-EI+ts or biphone EI+ts, @-l+gls, @-ts+E, C-a:+R, C-e+m, C-n=+s, E-C+n=, E-N+N=, E-S+n=, E-b+m=, E-k+N=, E-p+m=, E-x+t, E:-ah+t, E:-d+n=, E:-f+m=, E:-f+t, E:-g+N=, EI-C+k, EI-ts…

It would be interesting to get a complete list of all words that are missing. Then I could train or remove these words efficiently.

How I am developing Ralf's German dictionary version 0.1.9.9:

1. Improve phoneme quality:

<xsl:when test="contains(lower-case(../grapheme), 'möglichkeit')"><xsl:value-of select="replace($sierra, 'mœglɪçkaɪ̯t','møːglɪçkaɪ̯t')"/></xsl:when>

2. Check whether the object dictionary contains møːglɪçkaɪ̯t instead of mœglɪçkaɪ̯t:

ubuntu@ubuntu-desktop:~$ saxonb-xslt -ext:on -s:'/home/ubuntu/Documents/201006/german-0.1.9.5/german-0.1.9.8.xml' -xsl:'/home/ubuntu/Documents/201005/german-0.1.9.4/improve-german.xsl' -o:'/home/ubuntu/Documents/201006/german-0.1.9.5/german-0.1.9.9.xml'

3. Improve phoneme quality:

<xsl:when test="contains(lower-case(../grapheme), 'benzin')"><xsl:value-of select="replace($sierra, 'bɛnt͡sɪn','bɛnt͡siːn')"/></xsl:when>

(more…)

How I am preparing Ralf's German dictionary version 0.1.9.8:

1. Add the rule to improve-german.xsl:

<xsl:when test="ends-with(../grapheme, 'zustand')
or ends-with (../grapheme, 'zustands')"><xsl:value-of select="replace($sierra, 't͡sʊstant', 't͡suːʃtant')"/></xsl:when>

2. Add the rule:

<xsl:when test="ends-with(../grapheme, 'ständen')"><xsl:value-of select="replace($sierra, 'stɛnd', 'ʃtɛnd')"/></xsl:when>

3. Add rule:

<xsl:when test="contains(../grapheme, 'spiel')"><xsl:value-of select="replace($sierra, 'spiːl', 'ʃpiːl')"/></xsl:when>

4. Add rule:

<xsl:when test="contains(../grapheme, 'angebot')"><xsl:value-of select="replace($sierra, 'aŋeːboːt', 'aŋgeːboːt')"/></xsl:when>

5. Add rule:

<xsl:when test="contains(lower-case(../grapheme), 'stöcke')"><xsl:value-of select="replace($sierra, 'stœk', 'ʃtœk')"/></xsl:when>

6. I want only one terminal per role attribute, not several terminals. This rule removes the other terminals except the first terminal:

<xsl:when test="contains($terminal, ' ')"><xsl:sequence select="substring-before($terminal, ' ')"/></xsl:when>

7. Add rule:

<xsl:when test="starts-with(../grapheme, 'vorz')
and starts-with($sierra, 'foːʀt͡s')"><xsl:value-of select="replace($sierra, 'foːʀt͡s', 'foːɐ̯t͡s')"/></xsl:when>

8. Add additional phoneme element:

<xsl:when test="contains(lower-case(../grapheme), 'angabe')"><xsl:value-of select="replace($sierra, 'aŋgaːb', 'angaːb')"/></xsl:when>

9. Improve phoneme quality:

<xsl:when test="contains(lower-case(../grapheme), 'graphik')"><xsl:value-of select="replace($sierra, 'gʀafiːk','gʀafɪk')"/></xsl:when>

10. Generate version 0.1.9.8 via Ubuntu terminal:

ubuntu@ubuntu-desktop:~$ saxonb-xslt -ext:on -s:'/home/ubuntu/Documents/201006/german-0.1.9.5/german-0.1.9.7.xml' -xsl:'/home/ubuntu/Documents/201005/german-0.1.9.4/improve-german.xsl' -o:'/home/ubuntu/Documents/201006/german-0.1.9.5/german-0.1.9.8.xml'

Some phoneme elements will be duplicated. This issue will be fixed later.

11. Download Ralf's German dictionary 0.1.9.8, and import it into simon. Suggestions for improvement are welcome.