Tony: a tool for melody transcription

So I've been looking around a little bit to get a feel how the normalisation should work.

Conceptually I think it's always going to look like this (sorry if this is obvious):

1. measure input gain
2. amplify so that new gain = x (where x is our chosen level)

Re 1: I think we should follow the Replay Gain guys, who calculate the gain of the original signal as the 95%ile of framewise (50ms) loudness measurements. This is apparently grounded in perception theory: http://wiki.hydrogenaudio.org/index.php?title=ReplayGain_specification
For simplicity we could just use the RMS on frames as our loudness measurement, so by taking the 95%ile of that we have our input gain.

Re 2: We need to decide how conservative we want to be on the output gain. If we do normalisation internally in Tony, then it's not a big issue, and we can adjust the floating point representation to look and sound right, e.g x = -10 dB, and abs values > 1 are fine. If we want to write a new file, which clips at abs values > 1, then we might need to be more conservative. x = -20dB seems to be what Brecht suggested, and in a few informal experiments on our singing data no clipping happened in that case.

A quick fix could be to simply normalise to maximum level, i.e. in Matlabby code:

x = x/max(abs(x));

That would, in many cases, be enough, and would introduce no clipping. (It might introduce some quantisation error, but I assume that's negligible.)

OK, as of ed9296a27a27 we normalise to max level == 1. Try it out.