Loading Events

« All Events

  • This event has passed.

A 1-D Search Method With Stable 1-Norm Solution for Linear Prediction

November 15, 2017 @ 2:00 pm - 3:00 pm

DATE:             Wednesday,  November 15, 2017

TIME:              2 pm

VENUE:          CSD 308

SPEAKER:      Jayesh M. K.  (EE11D040)

GUIDE:           C. S. Ramalingam

DC Members :

Dr. Devendra Jalihal (Chairperson)

Dr. Arun Pachai Kannu (M)

Dr. Chandra Sekhar.C (M)(CS)

Dr. Hema A Murthy (M) (CS)

Dr. Umesh.S (M)



The vocal tract is commonly modeled as an all-pole filter.  The filter parameters are typically estimated using the minimum mean-squared error (MMSE) criterion because it leads to a set of linear equations.  The principal drawback is that the estimated filter has peaks that are not at their true locations but biased towards the harmonics of the excitation.  The bias increases with increasing pitch, to the point that the modeling accuracy becomes poor for high-pitch voices.  On the other hand, if the 1-norm  error is minimized, the estimate is completely pitch-independent; the price paid is that stability of the filter is no longer guaranteed.  The existing methods for stabilization of the 1-norm solution are not satisfactory.  In this talk we explore a solution based on line-spectral frequency and propose an iterative method that provides a stable solution while minimizing the 1-norm.  We compare our results with those obtained from the Convex Optimization Toolbox (CVX) by using the spectral distortion and its peak as the metric.  We have used both synthetic and natural speech in our experiments, the latter  being 1 hour of voiced speech frames taken from the TIMIT database.



November 15, 2017
2:00 pm - 3:00 pm
Event Category:


porno izleredtube izleHD redtubethemeforest nulleddaily hd porn