Grammer for "ZERO"…"ONE HUNDRED AND FORTY FOUR"

[rochbu]

My word/phrase list includes the numbers 0…144.

Using the numbers in the array works. e.g. [“1″,”2″,…,”144”].
Logging reports them not being in the acoustic model file.

Using a straight array works ok. I imagine I can do some tuning by looking at multiple hypothesizes. e.g. [“ONE”,”TWO”,…,”ONE HUNDRED FORTY FOUR”].

Pondering using a grammar.

The first word expected being [“ONE”…”NINETEEN”,”TWENTY”…”NINETY”]
The next word is optional.
If it is there: [“ONE”…”NINE”,”HUNDRED”]

The next work is optional.
etc.

Couple of questions:
A) If the grammar forms a tree will it be more efficient? (versus all combos)
B) A number of the first words would not have following words assuming a restricted grammar (TWO ZERO not allowed for TWENTY)

e.g. TWO would always be by itself.

[Halle Winkler]

Hello,

Couple of questions:
A) If the grammar forms a tree will it be more efficient? (versus all combos)

What do you mean by efficient in this case? Do you mean will it result in a smaller amount of code in your app, or will it be more accurate for the user, or will it be faster to return a result, or…?

A number of the first words would not have following words assuming a restricted grammar (TWO ZERO not allowed for TWENTY)

I’m not quite seeing the question in this one, can you clarify?

[rochbu]

On efficient:
I was originally thinking faster to return a result.
If it is a tree structure then following branches implies less ‘comparisons’.

I may have assumed this would result in a more accurate result. which may not be the case. Lots to learn.

On grammer:

If one word in [one,two,…,nineteen,twenty,thirty,…,ninety] then done
If two words and first word is in [Twenty, Thirty, Forty, Fifty, Sixty, Seventy, Eighty, Ninety] and second word in [one,two,…,nine] then done
it three words and first word is [one] and second word in [oh,ten,twenty,thirty,forty] and
if second word was [oh] then third word in [one,…,nine] then done
else second word in [ten,..,forty] and third word in [one,…,nine]

which I may be able translate into the must one of, may one of, … now that I’ve written it out this far.
(and I’ve left out some options such as ‘one hundred and one’).

[Halle Winkler]

A stock OpenEars grammar will not be faster to return a result than a language model – finite state grammars are just slower to search. A RuleORama grammar may be as fast as a language model. Either type of grammar should return more correct results and exclude more non-matching utterances than a language model. I still don’t really understand the second question, sorry. If I were you I would just test both approaches so you can see the differences firsthand in your app. Remember to test on a device.

[Author]

Posts