A TTS is too chatty and "smart" and at most situations a builtin converter from acronims - abriviations to words is not accurate.

[gregjozk]

Hello,

I'm raising this issue, due to on the NVDA's site on Github there are more and more issues, which are firstly linked to TTS and a default TTS in NVDA is espeak NG or Onecore, which seems to inherit most code from espeak NG due to same mistakes that it is doing.

here is a problem: a synth treats both acronims and abriviations the same so "cm" is always "centimeter" but in most cases it is not correct.
there are many more cases, where this behaviour is incorrect. e.g.:
in slavic languages "you" in plural is "vi", but espeak it treats always as "roman 6".

also test, how is spelled "x64". in most cases it is pronounced as "ten sixty four".

possible solution: do not convert abriviations to words, because in different situations they have different meanings an leave tts to spell it or speak it as usual word if it is written with lowercase included.

thanks.

regards,
Jožef

[valdisvi]

About which language are you talking about?

[jaacoppi]

All TTS software pronounce words based on assumptions about the language and context of the sentence. The modern solution would be to pass meta information about the sentence to the TTS. One such solution is part of speech tagging (POS). Since the origins of espeak NG are in 1995, these modern solutions are not used. Instead, all choices are rule based. It would be great to give user more configuration options to choose the preferred behavior. That is a work in progress with LoadVoice() and LoadConfig() functions. Unfortunately the progress is slow because the original code wasn't designed to support all these languages and situations. Currently, the only option for the user is to edit the config files and rebuild the code. That's probably too difficult for most users.

[gregjozk]

It is general observation, but seen mostly during english and slovenian use, which have been used daily. 2021-09-14 11:15 GMT+02.00, jaacoppi ***@***.***>:

…

All TTS software pronounce words based on assumptions about the language and context of the sentence. The modern solution would be to pass meta information about the sentence to the TTS. One such solution is part of speech tagging (POS). Since the origins of espeak NG are in 1995, these modern solutions are not used. Instead, all choices are rule based. It would be great to give user more configuration options to choose the preferred behavior. That is a work in progress with LoadVoice() and LoadConfig() functions. Unfortunately the progress is slow because the original code wasn't designed to support all these languages and situations. Currently, the only option for the user is to edit the config files and rebuild the code. That's probably too difficult for most users. -- You are receiving this because you authored the thread. Reply to this email directly or view it on GitHub: #1014 (comment)

[valdisvi]

Because of reasons @gregjozk mentioned, particularly for Latvian, there are almost no expansion of abbreviations except very few unique ones, because they are different in different domains. Therefore, for most of cases, there is no acute need to adjust behavior dynamically, unneeded expansions of abbreviations just need to be removed from rules.

[valdisvi]

I can't reproduce described behavior in development version of eSpeak NG for Slovenian:

espeak-ng -x -vsl "vi x64 10 cm"
S'e:st 'iks St'i:RiinS'e:zddEsEd dEs'e:t ts,@m'@

For English only vi is pronounced as "roman six":

espeak-ng -x -ven "vi x64 10 cm"
r,oUm@n_ s'Iks 'Eks s'Iksti f'o@ t'En s,i:;'Em

[gregjozk]

Hello in english "vi" is spoken as roman six exactly an in slovenian it is spoken as "šest" instead of "vi" meaning "you". tested with latest NVDA using espeak NG. so to clarify, "vi" should be spoken in slovenian as "vi" not as "šesst or rimska šest" etc, because "vi" in slavic languages means "you" for plural. thanks. 2021-09-15 21:27 GMT+02.00, Valdis Vitolins ***@***.***>:

…

I can't reproduce described behavior in development version of eSpeak NG for Slovenian: ``` espeak-ng -x -vsl "vi x64 10 cm" S'e:st 'iks St'i:RiinS'e:zddEsEd dEs'e:t ***@***.***'@ ``` For English only `vi` is pronounced as "roman six": ``` espeak-ng -x -ven "vi x64 10 cm" ***@***.***_ s'Iks 'Eks s'Iksti f'o@ t'En s,i:;'Em ``` -- You are receiving this because you were mentioned. Reply to this email directly or view it on GitHub: #1014 (comment)

[jaacoppi]

roman numbers are set in tr_languages.c with the keywords NUM_ROMAN, NUM_ROMAN_CAPITALS and NUM_ROMAN_ORDINAL. They can alse be set in the language files like espeak-ng-data/lang/zls/sl with the keyword 'numbers', but the documentation is poor. There are many other tickets which discuss completely removing the roman number support for all languages since using them is often the wrong assumption. No decision has been made yet. Should we do it now? My goal is to make LoadConfig() call LoadVoice() in such a way that the settings could be updated runtime instead of compilation time. That way the roman numbers and many other options could be off by default and set on by users if they need them. Unfortunately I don't have time right now.