Recognizing auditory signals

[Seanny123]

In "Long Term Memory for Noise: Evidence of Robust Encoding of Very Short Temporal Acoustic Patterns", the subject were asked to discriminate 1s cyclic noises (where the two halves are equivalent) or 1s random noises. Four weeks later, target cyclic noises that were shifted or chopped+shuffled were presented in an another discrimination task. The discrimination for intact cyclic noise was significantly better than for novel cyclic noise presented four weeks later.

This should be possible to model using rolling-window network.

To be more specific, the noise is just Gaussian noise:

The sound stimuli were sequences of normally-distributed, 16-bit pseudo-random numbers with a zero mean, which were played at a sampling frequency of 44.1 KHz. To ensure that the sounds are different every time, we reset the seed of the pseudorandom number generator of MATLAB on every trial.

I was talking with @tcstewar and he told me that for pre-processing, we could just select a set of frequency bands and encode their power spectrum into a set of dimensions. These dimensions could then be fed into the aforementioned rolling-window network for classification.

[arvoelke]

For sake of posterity, one of @celiasmith's students in SYDE 556/750 looked into this paper using the rolling window for his course project. A binary classification is learned online using PES and it is concluded that this captures the general trend.