Statistical Modeling and Estimation

The PhlSEM synthesis algorithm was described in detail in Chapter 13, but weTl briefly recap it here. The basic idea is to model the likelihood that a 

So far we have a simple synthesis model, but we still lack complete techniques for actually analyzing walking sounds, specifically to determine N—the Poisson probability constant. N is estimated by inspecting a high frequency band (5.5-11 kHz) of the whitened footstep sounds. A Daubechies four-wavelet filterbank is used to split the signal into subbands, and these subbands are rectified (absolute value). As can be seen by inspecting the top 

a threshold value is computed from the maximum peak and average values of the rectified subband signal, and the number of peaks per second over the threshold are counted. This one estimate of N (the density of collisions), tends to miss collisions as the probability N increases. To correct for this, the actual N estimate is calculated to be  

Using the PhlSEM model, 1050 soundfiles were synthesized using a simple raised cosine excitation envelope. System parameters were  

---