1 Frequencial Masking

The HAS (Human Auditory System) has a finite frequency resolution, which basically means that a weaker audio signal (maskee) becomes inaudible in the presence of (is masked by) a louder audio signal (masker), when they are close enough [1], in the frequency domain (and obviously in time, i.e., in the same chunk). When this happens, the subband [2] in which the maskee signal is placed can be quantized more severely without perceiving the quantization noise in the maskee subband (see Figure 1).

2 Dynamic computation of the quantization step sizes

Taking advantage of simultaneous masking involves adapting the quantizer step size \(\Delta _b\) (which is different in each subband \(b\)) to the energies of the subbands, and this can be difficult to do in real-time. Fortunately, modern computers can run multiple processes in parallel, and the relative energy between subbands doesn’t usually change drastically between chunks. Therefore, it’s possible to run the task of recalculating \(\Delta _b\) in a new thread/process, at a frequency that your machine allows.

Take into consideration also that the simultaneous masking effect can be computed considering a sliding window of adjacent chunks. This should provide a smoother transition between chunks.

3 Requirements of the implementation

Your implementation should take the following aspects into account:

1. It should not cause interruptions in audio playback.
2. It should maximize, for a given bit-rate, the PSNR of the reconstructions.

4 Deliverables

A new InterCom layer written in a Python module named simultaneous_masking.py. This file should be generated, explained and evaluated in a notebook named simultaneous_masking.ipynb. Consider only the MDCT case.

5 Resources