Here're some test result, I tested in a very quiet room with MacBook Air, volume 81%, pronounced by Samantha, American English accent (https://developer.nuance.com/public/index.php?task=supportedLanguages).
I tested three sentences, 10 times for each one, and I found it's really hard to fully recognize sentence which is more than 4 words.
what's your name
- bless your name
- last year they
- what's your name
- let's see any
- what's your name
- what's your name
- what's your name
- what's your name
- what's your name
- what's your name
what's the weather like tomorrow?
- what's the weather lie
- what's the weather
- what's the website law
- what's the weather lie
- what's the west la
- what's the weather like
- what's the weather
- what's the weather though
- what's the weather though
- what's the weather
show me your passport
- show me your past
- show your path
- show me your cat
- show your path
- so you're
- some you asshole
- show me your path
- so you're
- show me your passport
- show me a pact
Last edit: Capt.Michael 2018-04-27
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
I've read tutorial and faq before, didn't find particular way to improve my situation.
I can't create acousitc and language models by myself, because I can't provide such amount of data. The only way is to use data downloaded from CMU on sourceforge.net.
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
I suggest you to give up and use Google API. Efficient speech recognition on mobile phone is still a huge research problem. There are no easy ways. You have to create another language model, tune the beams on a test set and probably use more modern neural network algorithms. We have some information on wiki page
Hi buddy,
I set "-allphone_ci" to true, as it runs on my low performance Android device (Quad-core 1.44GHz, 1G RAM).
But accuracy of Ngram searching is still not good, even not reach the leavel as usable.
Is there any other parameters could improve accuracy on low device?
I use cmusphinx-en-us-ptm-5.2 + cmudict-en-us.dict + en-70k-0.2-pruned.lm.bin. This combaination has a good accuracy on my Samsung S8.
Last edit: Capt.Michael 2018-05-02
Here're some test result, I tested in a very quiet room with MacBook Air, volume 81%, pronounced by Samantha, American English accent (https://developer.nuance.com/public/index.php?task=supportedLanguages).
I tested three sentences, 10 times for each one, and I found it's really hard to fully recognize sentence which is more than 4 words.
what's your name
- bless your name
- last year they
- what's your name
- let's see any
- what's your name
- what's your name
- what's your name
- what's your name
- what's your name
- what's your name
what's the weather like tomorrow?
- what's the weather lie
- what's the weather
- what's the website law
- what's the weather lie
- what's the west la
- what's the weather like
- what's the weather
- what's the weather though
- what's the weather though
- what's the weather
show me your passport
- show me your past
- show your path
- show me your cat
- show your path
- so you're
- some you asshole
- show me your path
- so you're
- show me your passport
- show me a pact
Last edit: Capt.Michael 2018-04-27
en-70k-0.1-pruned.lm.bin + cmudict-en-us.dict is almost the same accuracy.
Last edit: Capt.Michael 2018-04-27
I've read tutorial and faq before, didn't find particular way to improve my situation.
I can't create acousitc and language models by myself, because I can't provide such amount of data. The only way is to use data downloaded from CMU on sourceforge.net.
I do appreciate any suggestion!
Last edit: Capt.Michael 2018-04-28
I suggest you to give up and use Google API. Efficient speech recognition on mobile phone is still a huge research problem. There are no easy ways. You have to create another language model, tune the beams on a test set and probably use more modern neural network algorithms. We have some information on wiki page
https://cmusphinx.github.io/wiki/pocketsphinxhandhelds/