[Resolved] Not detecting correct letter / number

[foobar8675]

Hello Halle,

I’m trying to use OE to recognize numbers and the alphabet. I replaced the words in the OE example to the following
NSArray *firstLanguageArray = @[
@”ONE”,
@”TWO”,
@”THREE”,
@”FOUR”,
@”FIVE”,
@”SIX”,
@”SEVEN”,
@”EIGHT”,
@”NINE”,
@”A”,
@”B”,
@”C”,
@”D”,
@”E”,
@”F”,
@”G”,
@”H”,
@”I”,
@”JAY”,
@”KAY”,
@”ELLE”,
@”EM”,
@”EN”,
@”OH”,
@”PEE”,

@”QUE”,
@”ARE”,
@”S”,
@”TEE”,
@”YOU”,
@”VEE”,
@”DOUBLEYOU”,
@”EX”,
@”WHY”,
@”ZEE”,

];

and am noticing the incorrect result come up pretty frequently. Suggestions?

[Halle Winkler]

This is unfortunately a thing which gets bad results with speech recognition. Most of the words are a single syllable, many of those syllables rhyme with each other (three, e, d, g, p) or nearly rhyme (eight and a and j), and they are contextless – they can arrive in any order (in contrast to phrases of words where we can say that some combinations are more likely, or required by a rule, versus others). You can probably find several other discussions on these forums about this, but they mostly consist of me making this approximate statement.

You can fix the fact that the letters are spelled phonetically, since that will prevent them from being found in the lookup list and given multiple pronunciations if there are some, but the main way to improve would be a specification change of some kind (organize your app so that only a few of these items are in any given language model or grammar – you will still encounter the e/d/g/p issue, but it will be a bit reduced).

[foobar8675]

Thanks for responding Halle. Do you think rulesorama might work? I was thinking, for example, if a user spelled the name for one of 3 users :
Tom
Ben
Mike

the grammer could look like

ThisWillBeSaidOnce : @[
@{ OneOfTheseWillBeSaidOnce : @[@”T”, @”B”, @”M”]},
@{ OneOfTheseCanBeSaidOnce : @[@”O”, @”E”, @”I”]},
@{ OneOfTheseCanBeSaidOnce : @[@”M”, @”N”, @”K”]},
@{ OneOfTheseCanBeSaidOnce : @[@”E”]},

Or is it obvious to everyone but me that this would that fall apart with a larger number of names (say over 100)?

[Halle Winkler]

I would just get a corpus of 100/200/300 names and try it out.

Your ruleset should look like this I think:

OneOfTheseWillBeSaidOnce : @[
@{ ThisWillBeSaidOnce : @[@”T”, @”O”, @”M”]},
@{ ThisWillBeSaidOnce : @[@”B”, @”E”, @”N”]},
@{ ThisWillBeSaidOnce : @[@”M”, @”I”, @”K”,@”E”]},

etc. I think that will do the recognition you intend, and limit the permutations to a number vaguely in the same ballpark as your corpus size. Plus it’s pretty easy to automate the dynamic generation.

[Halle Winkler]

Take a look at OpenEars 2.01 out today, which has a fix which should improve accuracy in some cases.

[Author]

Posts