# Concepts / Algorithms for recognising words in the Time-Domain

I'm currently working on a project that involves in Speech and recognising Phonemes (in kind of real-time).

The current algorithm works by always listening for input from the Mic, and input that is above a specific threshold then this is recorded to a .wav file and sent to the server for processing. E.g.

listening...
listening...
listening.. (max > threshold)
recording..
recording..
etc..


What I want to do is therefore have it recording when a specific name has been said. Someone basically speaks to it.. I.e. (let's assume it's called iris) so the algorithm would therefore work like the following:

listening...
listening..
listening.. (max > threshold)
// determine if the input is "iris"
// if true:
begin recording the sentence
// else:
listening..


I want to do this in the time-domain, then, process the sentences in the frequency domain. I've read an article, and, a paper that describes this process using Total number of Zero-crossings (using a specific threshold) my questions is whether or not this would be a good measure? I only need to identify one word, that is "Iris" and then everything else can be done using algorithms such as: MFCC, HMM's etc..

Any ideas? :)

• Does this have to be speaker-independent (i.e. work for lots of different people) ? Can you "train" the system? Or does it have to work straight away without any other user interaction? – Peter K. Oct 7 '13 at 11:56
• @PeterK. Basically, it has to work with multiple people. But, I read the following: cs.dartmouth.edu/~dwagn/aiproj/speech.html and this guy uses Zero-crossings. But I think I have an alternative solution – Phorce Oct 7 '13 at 12:34