[Espeak-general] espeak-zh multiple tone 3s

[Silas S. B. <ss...@ca...>]

Dear eSpeak listers,

I originally advised jonsd that, in cases where
there are multiple tone 3s in Mandarin Chinese,
a tone 3 should be changed into a tone 2 if and
only if there are an ODD number of consecutive
tone 3's immediately after it.  Hence
ni3hao3 is spoken as ni2hao3,
hao3lao3ban3 as hao3lao2ban3,
ba3bao3xian3qu3xiao1 as ba2bao3xian2qu3xiao1.
This understanding was based on the rules set
out in beginners' Mandarin textbooks.

However, it recently came to my attention that
the phrase 把保险取消 (ba3 bao3xian3
qu3xiao1, "to cancel an insurance policy") is
spoken more like "ba3 bao2xian3 qu3xiao1" by every
Chinese native I asked, even though this leaves
two consecutive tone 3's (xian3 qu3).  One said
my naive rendition "ba2 bao3 xian2 qu3 xiao1"
is "definitely wrong" (not just "acceptable as
a variation", but "definitely wrong"), and 2
others independently said (without any
prompting on my part) that I had probably been
taught a rule to never put two tone 3's
together but I should now forget this rule.

A Wikipedia author recently had this to say on
http://en.wikipedia.org/wiki/Standard_Mandarin :

----------------------------------------------
Pronunciation also varies with context
according to the rules of tone sandhi. The most
prominent phenomenon of this kind is when there
are two third tones in immediate sequence, in
which case the first of them changes to a
rising tone, the second tone. In the
literature, this contour is often called
two-thirds tone or half-third tone, though
generally, in Standard Mandarin, the
"two-thirds tone" is the same as the second
tone. If there are three third tones in series,
the tone sandhi rules become more complex, and
depend on word boundaries, stress, and
dialectal variations.

Tone sandhi rules at a glance

   1. When there are two 3rd tones (˨˩˦) in
a row, the first syllable becomes 2nd tone
(˧˥), and the second syllable becomes
half-3rd tone (˨˩).

          ex: 老鼠 (lǎoshǔ) becomes
[lao˧˥ʂu˨˩]

   2. When there are three 3rd tones in a row,
things get more complicated.

          If the first word is two syllables,
and the second word is one syllable, the first
two syllables become 2nd tones, and the last
syllable stays 3rd tone:

              ex: 保管好 (bǎoguǎn hǎo)
becomes [pao˧˥kuan˧˥xao˨˩˦]

          If the first word is one syllable,
and the second word is two syllables, the first
syllable becomes half-3rd tone (˨˩), the
second syllable becomes 2nd tone, and the last
syllable stays 3rd tone:

              ex: 老保管 (lǎo bǎoguǎn)
becomes [lao˨˩pao˧˥kuan˨˩˦]

   3. When a 3rd tone is followed by a first,
second or fourth tone, or most neutral tone
syllables, it usually becomes a half-3rd tone.
The half-3rd tone is a tone that only falls but
does not rise.

          ex: 美妙 (měimiào) becomes
[mei˨˩miao˥˩]
----------------------------------------------

Rule 2 is a nightmare.  It means we have to
know where the word boundaries are, at least in
the case of two-syllable words that start or
end with tone 3.

When I made zh_listx, I did so on the
assumption that we don't really have to know
the word boundaries.  Getting word boundaries
right automatically is difficult - Wenlin
always throws back "segmentation ambiguities"
at the user, pinyinannotator.com often gets it
wrong, etc.  I thought it won't matter if in
most cases the location of the word break does
not affect the actual pronunciation, so I made
no attempt to get word breaks right (and in
fact a couple of the rules I've recently added,
的地 and 地的, deliberately cut across word
breaks, as a hack to try to ensure 地 is
pronounced correctly when used in literature
with unknown place names).

I can certainly modify my zh_listx generator
script so that it always includes two-syllable
words that start or end with tone-3 even if
these are not strictly needed as "exceptions".
Then the above rule 2 can be implemented by
looking at the spaces as well as the tones. 
BUT I'm not convinced that will fully solve the
problem, because the rules as stated above are
not complete.  For example, what if there are
two two-syllable words, each with two 3rd
tones? or a one-syllable word followed by two
two-syllable words, all with 3rd tones? or a
3-syllable word that not everyone agrees if and
how to break up?  The rules above do not say
anything about those cases, or how things are
prioritised when two or more cases overlap (as
they do in ba3 bao3xian3 qu3xiao1).

And it gets worse.  A current PhD candidate in
Cambridge who's investigating tone teaching got
2 professionals to record the phrase 好老板
(hao3 lao3ban3, "good boss") for one of her
experiments, and one said it a bit like hao2
lao2ban3 (but with the first tone 2 at a higher
starting and ending pitch than the second),
which is sort-of an application of rule 2 above,
while the other said hao3 lao2ban3, which is
applying my initial naive understanding of the
rules.  Perhaps he had read this and was trying
to be "right" even at the expense of naturalness?
(That's a problem you can have if you tell
people to record words very slowly for
learners, especially if they're from regions
that normally have accents and therefore might
over-correct in their attempts to be "standard".)

I do not have permission to attach the recordings
for the eSpeak development archive (I've only got
them because I gave her some technical assistance)
but I can try to summarise:

32m / 32f - 好老板 hao3 lao3ban3 (nice boss)
- as discussed above

33m / 33f - I think this is supposed to be 老捕懂
but I'm not sure.  It is 3 * 1-syllable words.
And the voices do exactly what they did on the
previous recording.  It is possible that their
pronunciation of this was influenced by the fact
that they had just pronounced something similar
(when reading lists it's easy to get stuck into
a pattern and pronounce words differently from
how you'd normally read them in non-list contexts).

34m / 34f - 保守党 bao3shou3dang3
(Conservative Party) - BOTH speakers change the
first 2 tone 3's in this case (the 1st 2 become
rising tones, with the 1st rising tone starting
and ending at a slightly higher pitch than the
2nd, in BOTH voices).  This is a 3-syllable
word that cannot really be broken down into 2+1
or 1+2, but if it CAN be broken down then I
think it would break into 2+1 保守 + 党). 
So I have no idea what is going on here wrt. the
sandhi rules.

35m / 35f - 小老虎 xiao3 lao3hu3 (little
tiger cub) - very similar to hao3 lao3ban3.
But note that some dictionaries list xiao3lao3hu3
as ONE WORD, so it's debatable whether it's
xiao3+lao3hu3 or xiao3lao3hu3.  See what I mean
about the problem with automatic segmentation.
I would argue that, in this case, IF it can be
broken down into 2 words THEN the break WOULD
go after the 1st syllable.  But how are we
going to get this kind of data for ALL 3+-syllable
words in the language?  In some cases I can get
it automatically by querying a dictionary for
both the 1st 2 and the last 2 syllables and see
if either of them make a word with pronunciation
that's similar enough to their use in the
3-syllable word, but I have a feeling that
there will be an awful lot of cases when BOTH
1+2 and 2+3 can make valid words - and then
what do we do?  Parse the definitions and
compare semantics?  Find free frequency data on
all words?  I was really hoping that basic
pronunciation would not be affected by
something this complicated.

36m / 36f 小老头 xiao3 lao3tou2 ("little old
man") - both voices change the xiao3 into xiao2
and leave the rest unchanged (no surprises there).

AND - the PhD candidate who is doing all this
(who is a native Chinese from Guilin) thinks
that the Wikipedia article is WRONG, although
she is not yet able to explain how to correct
it.  She says she's going to have to think very
seriously about this kind of thing in her thesis.

So, to summarize,

(1) The rules for 3 or more consecutive 3rd-tones
are more complicated than we thought.

(2) They might depend on where the word
boundaries are (which are difficult to find
accurately when the input is in Chinese hanzi).

(3) We still don't know for sure exactly what
the rules are, and we do have cases that are
interpreted differently by different natives.
But natives do think our basic interpretation
is "wrong", at least in the case of 把保险取消.

So what do we do now?  Wait for the thesis? 
Help her write it?  .....  我糊涂了 (I'm feeling muddled)

Silas

(The Chinese-takeaway kitchen worker who wanted
help cancelling her mobile phone insurance had
no idea what this was going to lead to!)

-- 
Silas S Brown http://people.pwf.cam.ac.uk/ssb22