Phase-compensation and time-alignment stage for gammatone filterbank

[hagenw]

@Hardcorehobel:

I created a new branch (Reconstruction) where I implemented the phase compensation and time-alignment stages to reconstruct the original time signal from Gammatone subband signals. There is a new Demo, termed DEMO_Gammatone_Reconstruction, which illustrates the processing. You should be able to see that the phase-aligned reconstruction achieves a higher quality in approximating the impulse response.

/cc @ningma97

[hagenw]

Just a question out of interest: does this work in a same way as the Hohmann reconstruction or is there another way to do it? How large is the time delay you have to add for reconstruction?

[Hardcorehobel]

Volker is using a delay of around 4 milli-seconds. This approach is fully compensating for the time delays, which can be in the range of up to 20 milli-seconds for the lowest channels. When working with the full signals, or large chunks, this approach will achieve the highest reconstruction quality. But of course the approach from Volker is interesting for real-time stuff with low-delay requirements.

[hagenw]

OK, I just wanted to know. I don't think that those 20 ms will be a problem. But we should remember that we have a difference. For example if a KS requests another feature from the AFE (beside the reconstructed signal) and is calculating some prediction value using both of them, the KS should now that it maybe has to compensate for a slightly time shift.

Could you please also create a branch Reconstruction in TWOEARS/documentation and update the section on the gammatone filterbank accordingly. And add at some point that it will introduce a delay of 20 ms.