Perturbations in the Hidden Space HBR 180898

To get an idea of the acoustic consequences of changing the parameters of our hidden representation, we show on this page spectrograms where one parameters varies at a time by some sinusoidal function of time.

As the orientation and scaling of the hidden space is arbitrary under the current constraints (i.e. any number of MLP/target solutions can be found which are related by a linear mapping), we have adjusted the MLP and targets to whiten the hidden space (i.e. make the targets have a unity covariance matrix).

Then to each parameter at a time, we apply a sinusoidal perturbation from the mean target vector. The surface acoustic form is shown below for all four parameters.

1. F1 decreases and F2 increases
2. F2 decreases and F1 decreases to a lesser extent.
3. The amplitude falls away - we need this to approximate silence.
4. The last parameter doesn't seem to have a larger effect, it could be that three parameters is sufficient for this data.