[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: Frequency to Mel Formula


Thanks again for your great explanations of this complicated stuff.

All that notwithstanding, I'm still poking around at why we have these two different mel scales, with breaks at 700 and 1000. So I got hold of Fant's book, which has Baranek's data table in it, and plotted up some comparisons.

See http://dicklyon.com/tech/Hearing/Mel-like_scales.svg

The "Mel 1000" curve comes pretty close to the Baranek table data up through about 4 kHz, then diverges far from it above that. The "Mel 700" curve misses pretty badly around 2-6 kHz, but fits better on average if you count the highest frequencies.

The "Umesh" curve, f / (0.741 + 0.00024*f), doesn't fit particularly well, but has a good shape, so I did a "fit" and got f / (0.759 + 0.000252*f).

I also did a mel-type fit, and found a broad optimum for the corner around 711.5 Hz (under the constraint that 1000 Hz maps to 1000, which I should probably have tried relaxing, but didn't).

Anyway, here's my theory: Fant fitted to the frequency range he cared about, which probably only went to 4 kHz or so. And then someone else probably did a fit to the same Baranek table over the whole range, and got the 700 number (the plot shows that the 711.5 point are pretty much right on the 700 curve). And that's why we see Baranek referenced so much, maybe?

I also looked at goodness of fit (sum squared error in mel space) including all the frequencies in the Fant/Baranek table. It turns out that the Umesh type fit has only 1/8 as much error as the mel-like fit, due to the Bark-like curvature at the high-frequency end.

So for people who like Baranek's table (assuming Fant has a true copy of it), the Umesh type function should be a win. But I don't think that function extends well to the larger log-like range that we find in the ERB and Greenwood type curves, which are the ones that make more sense in auditory-based applications.

That's my theory and I'm sticking to it.