[orca-list] Speech in general (Re: Capital, Capital, Capital)

[Willie Walker <William Walker Sun COM>]

Hi All:

I think it's important to take a step back and look at the overall 
problem we're facing.  We have the desire for:

* Verbalized punctuation and capitalization
* Verbalized characters (e.g., 'double you' for 'w')
* Verbalized key names (e.g., 'Left Shift')
* Phrase spelling (letter-by-letter and 'military style')
* Customized pronunciation
* Abbreviation expansion
* Homograph disambiguation (the 'live' in 'Where do you live'
  vs. 'I live in cave')
* Natural F0 contour and prosody handling
* Audio cues
* Progress callbacks (e.g., 'this word was just spoken')
* Voice/pitch changes
* Multilingual support
* Etc.

A lot of these can be done at the speech layer without the need for 
additional knowledge.  Things like voice/pitch changes based upon 
context (e.g., it's a link, it's a pushbutton, etc.) may still need to 
live at the screen reader level.  Locale knowledge may also need to live 
a little higher in the stack.

We have a variety of speech synthesis engines with limited 
standardization for any of the above, each of which does all or a subset 
of the above in different ways.  SSML is an interesting thing, but only 
a handful of engines really support it.  In addition, it really doesn't 
provide support for things like 'say-as=military_spelling'.

We also have at least a couple kinds of users: 1) those that don't 
necessarily care about the details of what the speech synthesis engine 
supports and just want all of the above to work, 2) those that are more 
aware of the TTS engine's capabilities and are willing to work with its 
limitations.  Many of these same users are willing to pay money for a 
high quality commercial engine and expect Orca to work perfectly with 
that engine.

This is a pretty complex problem.  The solution we're currently working 
with in Orca is that it handles a lot of the above.  Based upon the 
discussion, I think we're agreeing there should be some delegation to 
the lower layers.  With this delegation approach, if Orca can learn that 
the lower layers support a feature, it can delegate responsibility for 
that feature to the lower layer.  If a lower layer doesn't support it, 
then Orca needs to provide it.

The first difficult task is figuring out how to obtain the information 
to do the appropriate delegation.  There's no standardization across any 
of the engines.  Take, for example, obtaining locale information.  The 
programmatic representation of locale differs greatly from engine to engine.

Take, for example, verbalized punctuation.  The engines that support it 
have their own ideas of 'none, some, all'.  I'll guarantee you that 
delegating verbalized punctuation to the engine will result in at least 
one member of this list shouting angrily that some punctuation mark was 
spoken at the 'some' level with one engine, but not another.

As a way to move forward, I think we might need to fill out the desires 
above and see what can be done to address them.  The TTSAPI work does 
some of this, but I'll admit it was done before I had a better 
understanding of the screen reader problem: 
http://www.freebsoft.org/tts-api.

Will