After people work through the [cnmat:node/2714|CNMAT Spectral Tutorials], they tend to ask "how do I make my own models?" Right now, the state-of-the-art is fairly primitive. Expect a lot of trial and error, and maybe unsatisfactory results.

#1. Choose your sound

This is the most important step in getting a good result. Some sounds analyze better than others.

In the sinusoidal world, problematic sounds are those with a lot of reverberation and/or a large noise component. For instance, I've never gotten a good shakuhachi analysis; shakuhachi is very noisy and the dominant recording style is highly reverberant. On the other hand, I think that the saxophone analysis in the spectral tutorials came out really well. Transients are much more difficult for sinusoidal analysis than harmonic tones. Singing voice is easier than speaking voice.

For resonant analysis, analyze a sound that decays away to zero amplitude by the end of the sound file. This is important, otherwise the analysis will give you some non-decaying filter that will probably blow up.

#2. Choose your software

##Resonant Analysis

If you are analyzing for resonant synthesis, there is basically one choice: ResAn. The other method to obtaining a resonant model is to convert a frame from one of the sinusoidal analyzers into a resonant moddel.

##Sinusoidal Analysis
Different sounds work better with different analyzers, and each piece of software offers different affordances for tweaking and altering the resultant data. Both the sax and shakuhachi models in the Spectral Tutorials were made with AddAn.
Some interesting results are here: http://www.cnmat.berkeley.edu/SDIF/ICMC2000/

NOTE: you need to have the SDIF framework installed in order to load AddAn.

You can find it here:
http://sourceforge.net/projects/sdif/files/sdif/SDIF-3.11.2/SDIF-3.11.2-…
It must be placed in YourHD/Library/Frameworks/*

#3. Specific Analyzers

See the pages below for advice on using specific analysis software.

Attachments
Image icon resan-new-analysis.jpg