Re: a couple of minor espeak ng issues in latest master


Reece H. Dunn
 

On Mon, Feb 27, 2017 at 10:59 pm, kendell clark wrote:

hi
I'd love the different english accents idea. Would these be different
varients selectable in the usual way or different voices since varients
are basically voices now and voices are languages?

The language is the first part, e.g. 'en' for English. The rest of that will specify the accent. 'en-US' for a modern generalised American English accent, like how 'en-GB' is used for a modern southern Brittish English accent. The other accents would be things like 'en-US-x-south' for Southern American, 'en-US-x-usny' for a New York accent, etc.

Voice variants in espeak terminology are what I am describing as voices in espeak-ng. These voices will support different voices, such as the MBROLA voices or alternative voice data to the current espeak voice and its klatt variants.

I'm ok with software
sounding like thought and bought, however it might need ... opening up a
bit. A little more like ah, like in palm, but not quite. God that sucks,
but it's the best I can put it.

I have now mapped the CLOTH vowels to /O/ which uses a sound similar to PALM. Does this sound better? Also, should THOUGHT, etc. be a similar sound to this?

I've had people tell me they've wanted a
more southern espeak ng voice for a while, but I didn't have the
slightest clue how to do it.

It requires understanding the Southern American accent, and adding a new Southern American phoneme table to espeak/espeak-ng.

I don't have many testers, just a few, but
one of them expressed a lot of interested in improved klatt voices. I'm
interested in that as well but they seem to require a lot of work.

This is something on my long TODO list. I want to get the architectural changes to make it easier to use different phoneme tables with the same language files. Once that is done, I am going to make the MBROLA voices use their own phoneme tables to provide better MBROLA voice support. After that, I intend to look at creating better klatt voices.

Not
that I'm complaining at all, it's just a different phonetic model and I
don't understand it enough to fix it. I'm going through the rest of your
email now and will reply to it. Just had to answer here.

No problem. Your feedback and comments are always welcome. I just don't have time to do everything at once.


Cheers

Reece

Thanks
Kendell Clark


Reece H. Dunn wrote:


On Mon, Feb 27, 2017 at 09:27 pm, kendell clark wrote:

hi all
First of all reece, I've just updated to the latest espeak ng
master and
I love the new changes made to english US english. They fix issues I
thought would never get fixed, making espeak ng's US english sound
much
more ... well, american.

I'm glad you like the fixes. I'm working through changes to
fix/improve the English pronunciations.

There are a couple of minor issues I thought
I'd better report though. First, are words like on, software, etc.
They
sound more like awn, sawftware. I'm not sure if this was intentional,
but my testers were hoping it could be changed back without removing
your cloth vowel fix.

Do you/your testers want the American accent to have the COT-CLOTH
merger? That is, do you want PALM and LOT (COT) to have the same vowel
sound as THOUGHT and CLOTH? Does this also include words like ORANGE
and HORROR (but not BORROW and a few others)?

I am using and fixing my https://github.com/rhdunn/amepd project,
which is based on the CMU pronunciation dictionary for American
English. I'm also creating a similar dictionary for Received
Pronunciation, and potentially others.

I'm thinking of creating other accents for American English as well,
with different features. I'll probably start by creating an
en-US-x-genam accent that has things like the COT-CAUGHT distinction,
while en-US has them merged.

The second is more serious but I'm not sure of a
fix. I'm not sure if this is a new issue or one I'm just now
discovering, but abbreviations don't work if followed by an
apostrophe
and an s. Example, ibm is said as Ib@ms, and nsa's is said as
Ens0z. n
soz for those who don't understand espeak phonetics. I figured I'd
leave
it to you to fix since I don't have the slightest clue how other
than to
fix every single abbreviation manually and there's got to be a faster
way.

This is a problem with the code. It looks like espeak/espeak-ng is not
preserving/using the $abbrev flag once it has removed a suffix like
's. I'll investigate.

The last one isn't an issue, but a request that I'm sure you've
heard before. Would it be possible, when you have time of course,
to fix
espeak's speaking of times and dates? Example, espeak says 2:00 as
2 00
instead of 2 o'clock. I'm not sure how to fix this one because this
would result in 2:00:00 two minutes and zero seconds or two hours
exactly, to be said as 2 o'clock, unless you could somehow make
espeak
speak it only when followed by am or pm, but that would break
military
time. I'm nt sure. Last, and then I'll go away, are dates, like 1950.
Espeak ng says 19 hundred 50, when 19 50 would be nice. Again, not a
complaint at all, just a request when you ahve time. I remember you
saying it would require a lot of refactoring of how espeak ng handles
numbers, so this one's definitely low priority, at least I don't
want to
act as if it's urgent, but it would be nice eventually.

The main problem here is detecting when a date or time is used, or
whether it is something different like a ratio. For example, the odds
were 12:15. This will likely be looked at with the other
context-related issues.

I love the fixes
you've done, if I weren't already using espeak ng as my only synth
this
would've made me switch.

Thanks. I have more improvements that I am working on, it just takes
time to check and fix the pronunciations.

Can't wait for an android update.

I have the infrastructure in place to be able to support Android
updates with each release now, so it should be possible to update the
Android version more often.

Finally, one
issue I'm reporting on behalf of a french speaker who uses the french
voice for english words as well. He reports that instead of
falling back
on the english phoneme tables it uses it's own french rules for
english
words. I'm not sure what the right approach is here. If there is a
french maintainer, would you mind taking a look? He's reporting that
words such as ubuntu, kubuntu, xubuntu, etc aren't said properly in
french, whereas they've been fixed for a couple of years in US
english.

These can be added to the French rules/lists files like is done in
English.

Also, there's a bug where if there's a word with a u and an i in it,
like suit, espeak ng french pronounces it as "sit" ignoring the u.
I'd
think the approach might be to have espeak ng french use the english
rules but there might be a really good reason to use it's own
rules, and
I don't want to act as if I know better.

The word suit/suite is actually derived from French :). The change was
made in 79003ea0521193a23e31aab6a5d90a7b7b2caedb by Thomas Guillory.
I'll revert that change so it sounds correct.

Kind regards,

Reece

 

Join espeak-ng@groups.io to automatically receive all group messages.