I created a new topic because i can't find the good key-words to find my answers.
As the subject says, i need to understand how the recognition proccess can be quick.
I would like to create an application that able to validates if the user said all the words display on a screen in the right order, and i would like to get statistics about the speed of him.
In fact, i would like something like a typing test (ttp://10fastfingers.com/typing-test/english) for speaking.
My thoughts are that in my case the system knows in advance a limited number of words expected,so i think the processing can be very fast.
What can i get in terms of performance and reliability. (ie: if the user is very fast like 50 words in 10 seconds) ?
What are the best conditions to decoding all words ?
Thanks,
Anthony
ps: i hope my english is correct, if not the case, i'm sorry for that...
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
In fact, i would like something like a typing test (ttp://10fastfingers.com/typing-test/english) for speaking.
Poor listeners of your students ;)
What can i get in terms of performance and reliability. (ie: if the user is very fast like 50 words in 10 seconds) ?
Decoder can detect words quickly and reliably in keyword spotting mode. It is what you need. You can tune detection threshold for a reliable detection.
There is a problem with fast speech though, modern approach to ASR doesn't scale well to a fast speech, If the set of prompts you want to verify is limited you can try to play with the dictionary and the model to deal with that. Reduced forms must be added in the dictionary. Some work is ideally needed on state-tying tree too.
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
If i well understood, i have to do some research to improve recognition for very fast speeches.
Do you think i can use CMUShinx "as is" to do that on Android ?
Thanks
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Yes, see keyword spotting search in android demo. For the long phrases you need to split them on sub phrases of 4-5 syllables though and spot for every phrase individually.
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Hi all,
I created a new topic because i can't find the good key-words to find my answers.
As the subject says, i need to understand how the recognition proccess can be quick.
I would like to create an application that able to validates if the user said all the words display on a screen in the right order, and i would like to get statistics about the speed of him.
In fact, i would like something like a typing test (ttp://10fastfingers.com/typing-test/english) for speaking.
My thoughts are that in my case the system knows in advance a limited number of words expected,so i think the processing can be very fast.
What can i get in terms of performance and reliability. (ie: if the user is very fast like 50 words in 10 seconds) ?
What are the best conditions to decoding all words ?
Thanks,
Anthony
ps: i hope my english is correct, if not the case, i'm sorry for that...
Poor listeners of your students ;)
Decoder can detect words quickly and reliably in keyword spotting mode. It is what you need. You can tune detection threshold for a reliable detection.
There is a problem with fast speech though, modern approach to ASR doesn't scale well to a fast speech, If the set of prompts you want to verify is limited you can try to play with the dictionary and the model to deal with that. Reduced forms must be added in the dictionary. Some work is ideally needed on state-tying tree too.
Thanks for your prompt reply :)
If i well understood, i have to do some research to improve recognition for very fast speeches.
Do you think i can use CMUShinx "as is" to do that on Android ?
Thanks
Yes, see keyword spotting search in android demo. For the long phrases you need to split them on sub phrases of 4-5 syllables though and spot for every phrase individually.
Thanks you very much for your answers :)