View Single Post
Posts: 959 | Thanked: 3,427 times | Joined on Apr 2012
#64
Originally Posted by nodevel View Post
It's a good idea, although the perfect implementation would be like this:
Me:"Hey Saera, write a text for me"
S: "Who is the text for?"
Me: "Jack" (triggers contact search) or "789-456-123"
S: "Ok, please tell me the message for /Jack/"
Me: "Hey, Jack, How are you? Done."
S: "Would you like to send the text now?" (if Yes, the word /Done/ won't be included in the text, if No, then it will and Saera will continue listening."

Then Saera could pass the text to the command line as it is possible to send messages that way.


Anyway, thanks for the great work.
That's more or less how I imagined writing a text. However, three things in that that I am still trying to figure out:
  1. Pocketsphinx uses a pre-trained grammatical model. This model apparently assigns a very low probability to multiple numbers being used in sequence, so it never seems to recognize a phone number. Even saying ten of the most distinctively pronounced number (seven), it only recognized four sevens. This is something I need to work on with the voice model, but have been putting off until I have enough time to recompile the model (maybe Sunday).
  2. Also, Pocketsphinx is not very good with names. This could possibly be alleviated by running a phoneme search on all contacts once it's determined to be not a number.
  3. If the user is dictating a text, there needs to be some way to edit what they said. Ideally, this would also train the voice model. This is definately possible, but I need to learn more about pocketsphinx first.

I had a hard time making Pocketsphinx recognize what I say. I was SO LUCKY to get Saera tell me the current time.
Is it possible anyway to 'train' the pocketsphinx to recognize what we say? Much like the microsoft voice recognition training.
Thanks a lot Taixzo !! Great Idea !!

Anything is possible with the N900
I'm working on it.
 

The Following 4 Users Say Thank You to taixzo For This Useful Post: