Hi Ralph,

>And I am convinced that in the long-term XML related standards are the way

>to go.  Maybe not today, but in a few years. 

I agree. 

>I would like to submit much more speech samples (prompts) in the English

>and in the German language employing the SSML, even if there isn't any

>demand at the moment. 

Please note that SSML is only a markup language for directing what a text-to-speech engine says.  I don't think it used as a format for describing transcribed audio submitted for the creation of acoustic models. 

>I started to read the HTK book.  It is really not easy.

The HTK book is a difficult read... I have only read the first few chapters, and now only use it as a reference - I don't have the math skills to understand all the formulas and how they interact.  But if you look at HTK as a "black-box", and only focus on the minimum command set required to compile an acoustic model, then you can do quite a bit with trial and error - which essentially was my approach when I started out... :)

You might be interested in the W3C VoiceXML standard, which essentially merges subsets of the SSML, CCXML, SRGML specifications.   This doc: "Voice Browsers, Introduction" provides a good overview of how they all should work together. 

The jvoicexml project has implemented a working VoiceXML browser, which essentially provides a VoiceXML dialog manager front-end to Sphinx and Festival.  They might provide bindings to Asterisk (IP PBX).  Note that jvoicexml uses the JSAPI and JTAPI "standards" to accomplish this.  

Ken 

Hello Ken,

OK, maybe it is not SSML what I am looking for.  I will take a look into the Introduction and Overview of W3C Speech Interface Framework, thanks for the link.  I am looking for "a format for describing transcribed audio."  Perhaps it is VoiceXML, I am not sure at the moment, I will read about it.

When I submit to VoxForge, there is the readme-file with information about the gender, pronunciation dialect, microphone type, audio card, operating system, sampling rate, etc..  And these information that are currently part of the readme-file should become part of a VoiceXML file.  So my goal is to merge the information from the prompts-file and the readme-file into one single VoiceXML file.

Yes, HTK - of course it is possible to do some trial and error.  I did that, but I didn't succeed.  A few months ago, I tried to follow the instructions from the VoxForge tutorial about HTK.  But there were too many errors for me.  I solved one problem, and then the next problem occurred.  I stopped trying.  Maybe I will try it again.

Greetings, Ralf

Another good way and a popular nowdays method to get a dictionary is the following. You select a phoneset, build an LTS system that will generate variants and then use forced-alignment against the recording to check are pronuncations valid or not. This way you'll ensure automatically that you dictionary is correct.

Also you would probably be interested to look on Unilex dictionary available from CSTR to check how the modern dictionary looks.

Hi nsh,

>You select a phoneset, build an LTS system that will generate variants and

>then use forced-alignment against the recording to check are pronuncations

>valid or not.

Forgive my ignorance, but by "LTS" do you mean "Letter to Sound"?   If so, do you mean that for each letter in the alphabet for a target language, you create a table that contains the different sounds that the letter might have, then you create a dictionary that would have multiple alternate pronunciations of the same word.  Then you take a transcribed speech recording and let the speech recognizer figure out the correct phonemes for each word (using forced alignment), based on what it recognizes in the recording?

For example, if someone wants to create a dictionary for a new language, do you first start with a set of speech transcriptions for the target language (i.e. speech audio files with a transcription of the actual words spoken in a text file). 

Then create the letter-to-sound rules.   For example the word "house" in the VoxForgdict is pronounced as follows:

HOUSE           [HOUSE]         hh aw s 

If I were using your approach, first I would create a phone list like this (CMU's phone list in this case):

Phoneme Example Translation
        ------- ------- -----------
        AA      odd     AA D
        AE      at      AE T
        AH      hut     HH AH T
        AO      ought   AO T
        AW      cow     K AW
        AY      hide    HH AY D
        B       be      B IY
        CH      cheese  CH IY Z
        D       dee     D IY
        DH      thee    DH IY
        EH      Ed      EH D
        ER      hurt    HH ER T
        EY      ate     EY T
        F       fee     F IY
        G       green   G R IY N
        HH      he      HH IY
        IH      it      IH T
        IY      eat     IY T
        JH      gee     JH IY
        K       key     K IY
        L       lee     L IY
        M       me      M IY
        N       knee    N IY
        NG      ping    P IH NG
        OW      oat     OW T
        OY      toy     T OY
        P       pee     P IY
        R       read    R IY D
        S       sea     S IY
        SH      she     SH IY
        T       tea     T IY
        TH      theta   TH EY T AH
        UH      hood    HH UH D
        UW      two     T UW
        V       vee     V IY
        W       we      W IY
        Y       yield   Y IY L D
        Z       zee     Z IY
        ZH      seizure S IY ZH ER 

I would then create a set of letter-to-phone rules as follows (phonemes converted to lower case for easier reading): 

H hh

O  ow, oy, uw

U uh 

S z

E iy

Then create rules for letter combinations to sounds (only for such letter combinations that have a unique sound in the target language):

HO hh aa, hh uh,hh ow

OU aw

US  ax s,

SE  s

Then generate all the possible pronunciations for the word "house":

HOUSE hh ow uh z iy

HOUSE hh oy uh z iy

HOUSE hh uw uh z iy

HOUSE hh aa uh z iy

HOUSE hh uh uh z iy

... 

And then use the forced alignment feature of a speech recognition engine(like Sphinx, HTK, ...) to look the text of a particular recording (in this case of the single word "house"), and see what phonemes it identifies as the most likely used in the recording (HTK format in this example):

    0 9400000 sil -5373.277832 SENT-END
    9400000 10400000 hh -750.756897 HOUSE
    10400000 11300000 aw -659.823364
    11300000 12900000 s -962.888245
    12900000 13300000 sil -238.437622 SENT-END

Which can then be input into a script to create the final correct pronunciation to the word "house":

 HOUSE           [HOUSE]         hh aw s

Ken