Virtual Sound: Audio Synthesis Tools / News: Recent posts

I have now added a VibratingString class for simulating string and wind instruments. Example usage is provided in VSound.c.

Submitted a small bug-fix that removes a couple of files that were stale and which prevented compilation.

This also includes a couple of new files for text statistics.

Moving to 0.2.x, a lot of effort has been going into the machine learning side of things, not only because of my research, but because I think it is a cool way to generate sounds. There are also a couple of effects that you may find useful and some code bloat like the cryptoanalysis tool.

I have now started consolidating the statistical learning classes [by starting to build common interfaces and so on], and added an HMM class. There are a lot of examples to keep you busy. I also added some reinforcement learning algorithms and test benches. ... read more

Updated the CVS repository with more unstable new features. However I have become distracted and focused more on the nonlinear prediction side of things.

Changelog:
2004-03-31 Christos Dimitrakakis <dimitrak@idiap.ch>

* BioNeuron.c, BioNeuron.h, DiscreteSignal.c, DiscreteSignal.h, DiscreteSpace.h, Distribution.c, Distribution.h, EntropyEstimates.c, EntropyEstimates.h, LockPitch.c, LockPitch.h, MIEstimate.c, MIEstimate.h, PredictiveStateRepresentation.c, PredictiveStateRepresentation.h, TimeStamp.c, accounting.c, sequence_targets2HTK.c, test_Distribution.c, test_EntropyEstimate.c, test_MIEstimate.c, test_PSR.c, test_bioneuron.c, test_compression.c, test_filter.c, test_mackeyglass.c, torch2bin.c, vsound2txt.c:
New file.... read more

A haphazardly put-together version has been placed on CVS. Hope it works. I added a lot of stuff for processing speech data and for computing predictive features. Also included is an ugly calculation of spectro-temporal responses of predictors.

This release adds functionality for supporting filter networks.

It turned out that it might be useful to have some kind of general
class for arbitrating IO in filter networks. While in the
non-recursive network case you can get good performance out of having
large feed-forward buffers, in the recursive case you don't need so
large buffers since you can only write data one sample at a
time. Since there are some problems with multi-rate filters and sample
windows, one would like to have asynchronous IO of some kind. So, the
(not tested) SIMOBuffer and IOPort take care of this functionality.... read more

OK, I have finally found the time to add the neural network predictive modelling code to vsound. There are quite a few different example main files that use them. I am not sure if they should go in the real actual application, as they need a lot of parameter fine-tuning in order to simulate sounds sufficiently well. Currently supported are ANNs with sigmoid or RBF units, with or without hidden control, and with normal or clustered predictors with varying delay, embedding dimension and lookahead.

No more features to be added in the first abstraction layer of effects and sample handing until the second layer (that handles multi-channel time-stamped samples) is finished.

Those that have downloaded dev-0.1.1 will have problems compiling, but I guess the can get around them. LITTLE_ENDIAN is redefined in SampleBuffer.h and it is useless anyway right now. Also, there is an unimplemented function in psr.c, LowPassFilter() - Remove it.

There will be another update over the weekend, the last for this year.