First time here? Check out the FAQ!

What are some possibilities for generating materials from batch analysis of audio files?

0 votes
More specifically, I'm looking for methods of batch audio file analysis that result in some kind of segmentation and ordering of the results, according to some gradient described by a parameter range or more multi-dimensional description involving several parameters.

update:  I've since found the link to the video tutorial on using the Batch Harmonic Analysis tool, listed in the User Guide as:   
In the video it mentions a pdf file about batch processing that used to be accessible from the help menu in Kyma X.  If that is still available and relevant, that might be a great help. Also, I just saw the section in the "Kyma X Revealed" document named "Writing Your Own Tools", within "Part 3: Advanced Topics".   Would that still be applicable for use within Kyma 7?  If so, I think I might need to spend some more time now with that older Kyma X manual..
asked Aug 4, 2016 in Using Kyma by thom-jordan (Practitioner) (670 points)
edited Aug 4, 2016 by thom-jordan

1 Answer

0 votes

Yes, the "Writing your own Tools" section in Kyma X Revealed is still relevant.  There are also some examples you could look at in the Xtra Tools folder inside the Kyma 7 folder.

Could you give a specific example of how you might want to sort or segment the files (on which combinations of parameters or features)?

Were you planning to be at KISS2016?

answered Aug 5, 2016 by ssc (Savant) (114,570 points)
I'd like to segment material with audible rhythmic transients always on an onset, or ambient materials at any relative location, then more specifically from there, perhaps at some point where novelty increases, based on analysis routines either performed in Kyma (once I learn if/how), or externally by providing the onset locations to segment at via OSC.  

I'd then want to sort all the segments into N number of ordered collections, where N is the number of individual analysis features (spectral centroid, skew, roll-off, kurtosis etc. or whatever's available or *can be developed*), and then refer to the results of each collection to look for segments that have desired characteristics in any number of dimensions, both for experimental use on the fly, and also perhaps to match model vectors in a self-organizing-map, where the results of the analysis-orderings might be sent back via OSC to an external process, which could then find the K nearest neighbors to a model vector and inform Kyma of the results, or alternatively (preferred), the process could query Kyma to find the kNNs if I can figure out how to write the required functionality in Smalltalk/Capytalk.

I very much wanted to attend KISS2016, but I decided to wait to see if next years event is within the USA.

Between where I'm at now and the main goal stated above, I can see many practical uses along the way while trying to develop something like the above. I use many algorithmic methods for event generation that are along the lines of what was originally called "total serialism", methods which seem to work amazingly well when accessing materials which are ordered in a certain way. Many times finding the most preferable ordering of materials constitutes a large part of the composition or performance preparation.

**  Robust versions of any low-level analysis features that I'd want to use are described in detail in "Audio Content Analysis" by Dr. Alexander Lerch of GTCMT:

...along with example implementations for each feature included in a accompanying code library for MATLAB.

If Smalltalk/Capytalk can be used to develop low-level feature metrics based on these algorithms for direct use within Kyma sounds, scripts, etc., then I can see myself starting some work on this right away, to develop a few choice features to begin with.
I hope you don't mind some advertisement.. In the NeverEngine Labs we have two packages that might be very interesting:

ROMTools: I wrote a tool that analyses a sample for transients and puts markers in he sample. Using Kyma's or our NE Labs MultiSample possibilities you can then playback the segments as you wish while keeping the file as a whole instead of splicing it into lots of small files.

SPC SR Vol. I: In there are some features I developed and encapsulated for easy usage like spectral centroid, brightness,... I want to do more of those in the future, like you I'm very interested in collecting meta data about samples for compositional use ;)

About Tools: You can do all those features using a Tool and smalltalk, but in my opinion it is more interesting (though ore challenging) to develop real-time algorithms so those features can be used live ;)