HQSYN16

After careful analysis and consideration, we came to the conclusion that a more detailed context-based penalization system (i.e., one going beyond nasalization and labialization contexts) is not realistic. We therefore propose only minor changes to the current system.

1. It is not clear why T (voiceless palatal plosive [c]) is not included in the nasalization matrix. We recommend that it is included.

2. Also for the nasalization matrix, we recommend that the following restriction - when examining diphone [ * ][ o,O,u,U,y,i,I,e,E,a,A ], disable the use of its l-context according to the following table - is only retained for obstruents (i.e., that it is dropped from the non-nasal sonorants [j, l, r]).

I am adding the list of penalties penalties.txt to be used during the experiment. It is quite large, since it is auto-generated from the test script.

But to avoid wrong conclusions, could you please look at it (briefly, just pick the most problematic items) and confirm that the penalties are ok?

• 10000 - disable. It is such a huge cost that this unit should never be selected (if there are other more appropriate variants)
• 10 - penalty. Another unit should be preferred, if possible, but the unit can be used e.g. in case when a concatenation cost to other units is too high
• 0 - no penalty

The lines in form have/want-[unit]: u/a-[d*] -> 10.0 means: penalise the use of unit with left context [u] when [a] is required for any diphone with left phone [d]
The lines in form +want/have: [*A]+C/J -> 1000000.0 means: disable the use of unit with right context [ň] when [č] is required for any diphone with right phone [á]

• diphone [ * ][ o,O,u,U,y,i,I,e,E,a,A ], disable the use of its l-context ... - should it be: disable the use of its r-context ...?

The penalties.txt uses left and right contexts. I will re-generate the penalties table if there should only be left context in both cases.

Outdated...