[rl1987]

Okay. Thank you.

[rl1987]

What kind of work needs to be done to make OpenEars recognize words spoken in non-American dialect? As far as I know, PocketSphinx is a general purpose speech recognition engine and could be used to perform speech recognition in multiple languages. Do I need to have recordings of Irish people pronouncing words of interest and perform model training to generate models that correspond specifically to Irish dialect of English? What are the specifics of doing this?

[rl1987]

Our application is intended for users who live in Ireland and speak the corresponding dialect of English. We want OpenEars to be able to recognize words spoken with this kind of dialect. Some of the testers belong to this set of intended users. We, the developers, are from East Europe, but we can speak English fairly well, although we do have our accents.

An Irish female tester reported that accuracy was as low as 3% when she was holding iPhone as one normally would. When she increased the distance between herself and device, the accuracy got better.

I have uncommented [OpenEarsLogging startOpenEarsLogging]; and tried generating the language model again.

When OpenEars generates the model, I am getting a fair amount of warnings like this:

2012-10-22 16:41:09.436 OpenEarsSampleApp[84775:11f03] The word LOCKSMITHS was not found in the dictionary /Users/rimantasl/Library/Application Support/iPhone Simulator/6.0/Applications/9309879C-9CFD-48C1-8A12-305A6EC7FDA5/OpenEarsSampleApp.app/cmu07a.dic.
2012-10-22 16:41:09.437 OpenEarsSampleApp[84775:11f03] Now using the fallback method to look up the word LOCKSMITHS
2012-10-22 16:41:09.438 OpenEarsSampleApp[84775:11f03] Using convertGraphemes for the word or phrase LOCKSMITHS which doesn't appear in the dictionary
2012-10-22 16:41:09.440 OpenEarsSampleApp[84775:11f03] If this is happening more frequently than you would expect, the most likely cause for it is since you are using the default phonetic lookup dictionary is that your words are not in English or aren't dictionary words, or that you are submitting the words in lowercase when they need to be entirely written in uppercase.
2012-10-22 16:41:09.451 OpenEarsSampleApp[84775:11f03] The word LONGFORD was not found in the dictionary /Users/rimantasl/Library/Application Support/iPhone Simulator/6.0/Applications/9309879C-9CFD-48C1-8A12-305A6EC7FDA5/OpenEarsSampleApp.app/cmu07a.dic.
2012-10-22 16:41:09.452 OpenEarsSampleApp[84775:11f03] Now using the fallback method to look up the word LONGFORD
2012-10-22 16:41:09.452 OpenEarsSampleApp[84775:11f03] Using convertGraphemes for the word or phrase LONGFORD which doesn't appear in the dictionary


These warning are mostly related to Irish location names.

Besides, there has been these warnings just after running OpenEarsSampleApp in iPhone Simulator:

sih_add WARNING: repeated hashing of 'GAMES', older value will be overridden.
sih_add WARNING: repeated hashing of 'GAS', older value will be overridden.
sih_add WARNING: repeated hashing of 'LEITRIM', older value will be overridden.
sih_add WARNING: repeated hashing of 'LESSONS', older value will be overridden.
sih_add WARNING: repeated hashing of 'MEATH', older value will be overridden.
sih_add WARNING: repeated hashing of 'MONTESSORI', older value will be overridden.
sih_add WARNING: repeated hashing of 'MORTGAGE', older value will be overridden.
sih_add WARNING: repeated hashing of 'RENTAL', older value will be overridden.
sih_add WARNING: repeated hashing of 'SALON', older value will be overridden.
sih_add WARNING: repeated hashing of 'SEWER', older value will be overridden.
sih_add WARNING: repeated hashing of 'SUPPLIES', older value will be overridden.
sih_add WARNING: repeated hashing of 'TARMAC', older value will be overridden.
sih_add WARNING: repeated hashing of 'TARMACADAM', older value will be overridden.
sih_add WARNING: repeated hashing of 'TYRES', older value will be overridden.
sih_add WARNING: repeated hashing of 'USED', older value will be overridden.
sih_add WARNING: repeated hashing of 'VETINARY', older value will be overridden.
sih_add WARNING: repeated hashing of 'WINDSCREEN', older value will be overridden.


[rl1987]

We have entered single words (one word per line) into phrases1.txt file and used the following code in the modified OpenEarsSampleApp to generate the language model:


LanguageModelGenerator *languageModelGenerator =
[[LanguageModelGenerator alloc] init];

NSString *phrasesPath = [[NSBundle mainBundle] pathForResource:@"phrases1"
ofType:@"txt"];

NSError *error =
[languageModelGenerator generateLanguageModelFromTextFile:phrasesPath
withFilesNamed:@"OpenEarsDynamicGrammar"];

Should the generated OpenEarsDynamicGrammar.dic contain only the words that aren’t found in the CMU dictionary? As of now, it contains all the words from phrases1.txt.

At worst, we would like to have at least 60% accuracy (that is, 6 successes from 10 experiments). 80% and higher accuracy would be good enough.

[rl1987]

Language is English, closest to American dialect I believe, although neither of the speakers are native Americans. Both males and females were trying to test this feature. All of the speakers have adult sounding voices. Also, for the testing I have been trying to use say(1) program on Mac OS X and been getting the same kind of problems.

We have been testing the application on iPhone 3GS, 4th gen iPod Touch and iPhone 4. In all cases, internal microphone of a device was used, there were no external microphones. OpenEars version is the latest one, 1.2.2. None of the audio-related iOS APIs were used in the code we have written since speech recognition is the only audio-related feature that is available in the application.

It seems that most of the phrases in my language model is found in the pre-packaged cmua07.dic file, but there are exceptions.

[Author]

Posts