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.
speech enhancement using Kalman-- help with segmentation
Started by ●February 4, 2007
Reply by ●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 butmost>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 usingKalman>filters. I understand Kalman filtering, no problem with that. Theproblems>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 coursemy>limited undestanding is deceiving me). >3). For colored noise a Voice Activity Detector is being employed butmost>papers are omittig the specific mathematical detais. Can you favor mewith>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. > > >