hi i have gone through simon haykin and lim , oppenham books on adaptive filtering before starting the problem is not in cal. coefficient but which coeff should be used for speech recognition we can't use only last coeff for differentiation so which coeff should be used as signal is nonlinear coeff changes till last the coeff. gets settle for some time & then as statistic property changes coeff change as expected so we r not getting which coeff should be used or we have to use coeff after successive intervals & then use dynamic time warping for comparing thanks in advance ...........ajay On Fri, 25 Jan 2002 Ganesan Ramachandran wrote : > hi, > i think u misunderstood what i meant by online LMS.. > i apologize for not being clear... i guess u haven't > had a formal introduction to gradient descent and > LMS.. if u know them please skip the lines between C > language comments (/* & */) > > /* gradient descent > the adaptive algorithms such as LMS, RLS & their > variants belong to a category called gradient descent > algorithms. the filtering operation done by this > adaptive filter can be considered as trying to > approximate a function in the multidimensional signal > space spanned using the delayed versions of the input > signal as the co-ordinate axes. if the filtering is > done in a linear fashion (i.e., with an FIR), then the > error between the desired and the output of the system > has a parabolic shape in the space spanned by the co- > efficients of the filter (i.e., weights) . so, to > travel down the parabola and to reach the minimum > error, we use the information from the gradient on the > parabolic surface. that's y these are known as > gradient descent algorithms. if u have trouble > understanding this, i suggest that u read any standard > adaptive filtering books like simon haykin or > principe. > > LMS > > in LMS there are two ways to do it. 1. online , 2. > batch.. the difference is whether u want to update the > weights instantaneous error (the e in ur equation) or > u want to update with an error averaged over some > samples. let me explain the online mode as it is > easier to implement and understand. > > in online LMS, what u basically do is say if u have > a filter order N, then u take N samples from the input > signal, pass it thro an FIR, find the error between > output and the desired and then use the error to > update the weights of the FIR... then replace the > filter input with the input signal values one by one > and proceed. since u handle with only N samples at a > time, u don't have to worry about the stationarity of > the signal and all as N is much smaller compared to > the duration of stationarity. > */ > > if u feel it goes way above the head, i strongly > suggest that u take a look at a standard book on > adaptive DSP.. if u need any clarifications, feel free > to contact me. > hope this helps. > Ganesan. > > hello, > i am planning to calculate by off line i.e. taking > speech from wav files. > > initial filter coefficient r set w=0; > and all new coefficients are updated by > Wi+1=Wi + 2*mu*e*u; > u i/p array > e-instantanious error. > next successive window we use previous as initial > coeff. > but the error is not converging > window size is 23.2 msec was taken & step size was > 23.2/25 > if i try to move window continuosly to avoid sudden > change of statestical char. it takes too much time > error gives same shape as of i/p > what is solution.???? > ...ajay > On Thu, 24 Jan 2002 Ganesan Ramachandran wrote : > > hi, > > how do u plan to calculate the coefficients? thro > > online/batch - LMS gradient descent? in that case, > u'll > > be dealing only with samples of length of the filder > > order > ------------------------ Yahoo! Groups Sponsor > > _____________________________________ > Note: If you do a simple "reply" with your email client, > only the author of this message will receive your > answer. You need to do a "reply all" if you want your > answer to be distributed to the entire group. > > _____________________________________ > About this discussion group: > > To Join: > > To Post: > > To Leave: > > Archives: http://www.yahoogroups.com/group/matlab > > More DSP-Related Groups: http://www.dsprelated.com/group- > s.php3 > > ">http://docs.yahoo.com/info/terms/ |