Reply by 2pers January 29, 20082008-01-29
Hi,

Just few answers:
> 1). Why the stationary assumed frames being windowed >non-rectangularly before extracting AR parameters? > 2). Furtermore, the >segments are being overlapped! Why?
This is because we want to estimate the PSD using the Welch periodogram method ...
>3). For colored noise a Voice Activity Detector is being employed but
most
>papers are omittig the specific mathematical detais.
That's true. Indeed, this is often part of the internal tricks. Yet, you can find some inspiration for doing that looking into GSM AMR code. Rgds. 2pers
>Hi Group! >I am new to the net. Hope I dont disappoint U with naive questions. >Well, a couple of papers I have gon thru on enhancing speech using
Kalman
>filters. I understand Kalman filtering, no problem with that. The
problems
>actually are: 1). Why the stationary assumed frames being windowed >non-rectangularly before extracting AR parameters? 2). Furtermore, the >segments are being overlapped! Why? This (overlapping)has increased the >number of input (noisy speech) data over what has actually been there.
By
>simply Kalman filtering it we increase correspondingly the number of >output data (filtered speech samples).Is not this anamolous? (of course
my
>limited undestanding is deceiving me). >3). For colored noise a Voice Activity Detector is being employed but
most
>papers are omittig the specific mathematical detais. Can you favor me
with
>a self-contained VAD algorithm demonstrating doc? So that a Matlab >implementation can be based on that. > I shall be highly obliged if anyone of above querries are >favorably acknowledged. -- Kul Mohit. > > >
Reply by Mohit February 4, 20072007-02-04
Hi Group!
I am new to the net. Hope I dont disappoint U with naive questions.
Well, a couple of papers I have gon thru on enhancing speech using Kalman
filters. I understand Kalman filtering, no problem with that. The problems
actually are: 1). Why the stationary assumed frames being windowed
non-rectangularly before extracting AR parameters? 2). Furtermore, the
segments are being overlapped! Why? This (overlapping)has increased the
number of input (noisy speech) data over what has actually been there. By
simply Kalman filtering it we increase correspondingly the number of
output data (filtered speech samples).Is not this anamolous? (of course my
limited undestanding is deceiving me).
3). For colored noise a Voice Activity Detector is being employed but most
papers are omittig the specific mathematical detais. Can you favor me with
a self-contained VAD algorithm demonstrating doc? So that a Matlab
implementation can be based on that.
           I shall be highly obliged if anyone of above querries are
favorably acknowledged.                              -- Kul Mohit.