Chris: I'm not sure I have anything to add to my previous comments in earlier threads which cover a few makes of mike. Most mikes will do very well unless you are using a really low end mike which adds noise to the system.

Could you explain in a bit more detail what conditions you will be using the mike in, parameters expected, etc? Eg noisy environment, wired/wireless, what OS you will be using (some mikes are more linux friendly than others, etc) ...

I would be using it for desktop command and control and home automation control. It would be inconvenient (for the home automation part) anyway to have to wear a headset each time you speak a command.  Ideally the microphone should have good pickup from a practical distance (1 foot perhaps or greater).  The microphone does not have to be wireless, but that would certainly help, though not wireless like a bluetooth headset that I have to wear.  The operating system is linux.  The sound card could be internal SBLive, or external extigy.  I would say the environment is not very noisy.  Omnidirectional would probably be the best, but cost is also an issue (no more than 100 dollars) .  I might be asking for too much here, but hopefully there is a reasonably priced solution. 

Would it help recognition if I record my samples from 1 foot distance.  I am creating a speaker dependent acoustic model to use.

Thanks

Chris

The fact that you would rather not wear the mike places your choice in an area that I do not know very well. When it comes to distance from mike, I have found that the model can be very forgiving of changes in mikes and distance from mike if trained with that in mind. If you decide that you will be giving most of your commands within x feet of the mike as you move about, then train the model with prompts recorded at varying distances in the range 1 to x feet. You may need to be quite creative with your dialog manager to avoid the lights coming on when the doorbell rings.

I have been reading a little bit about these far field microphones and it seems that other than ambient noise, there is the problem of reverberation so that recognition rates will always be worse with them.  Is there some kind of software filter available that can preprocess the audio and pipe to julius? 

Thanks

Chris

>Is there some kind of software filter available that can preprocess the audio

>and pipe to julius? 

You might try "spectral substraction" (i.e. remove noise from the audio being fed into Julius).  I have not tried this myself though, so I am not sure how well it works.

See Julius manual:

      -sscalc
              Perform spectral subtraction using head part of each file.  With
              this  option,  Julius assume there are certain length of silence
              at each input file.  Valid only  for  rawfile  input.   Conflict
              with "-ssload".
 
       -sscalclen
              With  "-sscalc", specify the length of head part silence in mil-
              liseconds (default: 300)
 
       -ssload filename
              Perform spectral subtraction for speech  input  using  pre-esti-
              mated  noise spectrum from file.  The noise spectrum data should
              be computed beforehand by mkss.  Valid  for  all  speech  input.
              Conflict with "-sscalc".