How to customize pronunciation in ScanSoft MeiLing?

[ssb22]

After being a satisfied user of Neospeech Lily for some time, I thought I'd like to try the ScanSoft MeiLing voice as well because its sample sounded good (and Lily does have a few minor bugs that I thought might not be in MeiLing).

I bought MeiLing, but now I am having terrible trouble editing the pronunciation of words. TextAloud's pronunciation editor won't do Chinese, and I tried the SAPI markup e.g.

<PRON SYM="fa 1 song 4"/>

- this works (sort-of), but the result sounds really awful. For example:

"fa 1 song 4" sounds like "fa-so-yi"
"hao 4 ma 5" sounds like "hao-low-yi"
'ren 2 men 5' says 'ran-men-ge'
'na 4 ge 5 ren 2' sounds terrible
etc etc
(so far almost nothing I've tried sounds anywhere near right).

"ni 2 hao 3" sounds more like "nay-hao" than nihao (the voice gets this one right from the characters, so I *know* it can do it, but why can't it do it from my own supplied pronunciation).

The voice is pretty useless to me if I can't edit pronunciation, since there's often more than one way to read a character (depending on the meaning of the surrounding words) and computers aren't very good at getting it right. My Chinese learning materials have pronunciation written over each character, and I *need* to be able to put this pronunciation in so as to make sure the computer is giving me the same as the book.

Could you give me some pointers as to how I can customise MeiLing's pronunciation properly?

Thanks so much.

Silas

[ssb22]

I had a little more success by feeding Zhuyin to the MeiLing voice, but it still won't completely work.

For example I typed ㄓㄨˋ 一ㄋ (the zhuyin representation of "zhu4yin1") into TextAloud (no PRON markup is needed here). MeiLing said it correctly.

However, when I typed ㄛˇ ㄕˋ 一 ˙ㄍㄜ ㄐ一 ㄑ一ˋ ㄖㄣˊ (the zhuyin representation of "wo3 shi4 yi1ge5 ji1qi4ren2", "I am a robot") it didn't work. The "j" and the "q" sounds were completely lost, as was the "w" sound ("w" is not explicitly represented in zhuyin, but "j" and "q" is).

My Zhuyin symbols (and their encoding) is correct according to the following references (which just might all be mistaken, in which case I'd like to know what I should be typing instead) :

http://en.wikipedia.org/wiki/Zhuyin_table
http://research.chtsai.org/papers/pinyin-xref.html
http://www.yellowbridge.com/language/zhuyin.html
http://en.wikipedia.org/wiki/Zh%C3%B9y% ... BAh%C3%A0o

[ssb22]

Note that in the Unicode standard, Zhuyin is known by its informal term "bopomofu", and the code chart is at http://www.unicode.org/charts/PDF/U3100.pdf
I have now confirmed the encoding I was using against that chart, so it's definitely not a Zhuyin encoding error that's causing this. Meiling does not appear to be able to accept pronunciation instructions for "j", "q" and "x" words, which are very common in Chinese. Am I missing something?

[ssb22]

Oh wait a minute, I made a mistake with that chart. For the "I" sound, I should have been using "bopomofu letter I" (unicode 0x3127), instead of the "one" hanzi which seems to be used everywhere else. So the "I am a robot" phrase should have been typed as: ㄛˇ ㄕˋ ㄧ ˙ㄍㄜ ㄐㄧ ㄑㄧˋ ㄖㄣˊ.

That's an improvement ("j" and "q" now sound correctly), but it still has bugs:
* ji is said with a falling tone (4) when it should be high level tone (1), and there's a pause after it which shouldn't be there
* the w sound at the beginning is still missing
* the rising tone of the final "ren" is barely audible

MeiLing happens to be able to get the phrase "ji1qi4ren2" right when given the Chinese characters, so compare the pronunciation of 机器人 with that of its phonetic representation ㄐㄧ ㄑㄧˋ ㄖㄣˊ to see what I mean (they should be identical but they're not). Note that relying on Chinese characters is not a good general solution because there are plenty of times in language learning where you need to specify things by pronunciation instead.

[ssb22]

Deleting the tone marks has no effect, so it appears that MeiLing is ignoring these marks, even though they appear to be encoded correctly. Maybe MeiLing is expecting some other means of specifying the tone? I'm not sure what though.

PS: in Unicode it's spelled "bopomofo", not "bopomofu" as I said above. (Both spellings are used elsewhere.)