<< Chapter < Page | Chapter >> Page > |
In effort to improve the efficiency of this algorithm, we created an alternative called FAST Autocorrelation, which has yielded speed improvements in excess of 70%.
We exploit the nature of the signal, specifically the fact that if the signal was generated using a high sampling rate and if the windows are narrow enough, we can assume that the pitch will not vary drastically from window to window. Thus, we can begin calculating the r(s) function using values of s that correspond to areas near the previous minimum. This means that, if the previous window had a fundamental period of 156 samples, we begin calculating r(s) for s = 136. If we fail to find the minimal s in this area, we calculate further and further from the previous s until we find a minimum.
Also, we note that the first minimum (valued below the threshold) is always going to correspond to the fundamental frequency. Thus, we can calculate the difference equation dr(s)/ds as we generate r(s). Then, when we find the first minimum below threshold, we can stop calculating altogether and move on to the next window.
If we use only the second improvement, we usually cut down the range of s from 600 points to around 200. If we then couple in the first improvement, we wind up calculating r(s) for only about 20 values of s, which is a savings of (580) * (1200) = 700000 calculations per window. When the signal may consist of hundreds of windows, this improvement is substantial indeed.
As stated earlier, autocorrelation is also extremely expensive computationally. However, using the adaptive techniques described above, computation can be expedited and run in near-real time.
The HPS involves two steps: downsampling and multiplication. To downsample, we compressed the spectrum twice in each window by resampling: the first time, we compress the original spectrum by two and the second time, by three. Once this is completed, we multiply the three spectra together and find the frequency that corresponds to the peak (maximum value). This particular frequency represents the fundamental frequency of that particular window.
Another severe shortfall of the HPS method is that it its resolution is only as good as the length of the FFT used to calculate the spectrum. If we perform a short and fast FFT, we are limited in the number of discrete frequencies we can consider. In order to gain a higher resolution in our output (and therefore see less graininess in our pitch output), we need to take a longer FFT which requires more time.
Notification Switch
Would you like to follow the 'Ece 301 projects fall 2003' conversation and receive update notifications?