Hearing AI: Getting Started with Deep Learning for Audio on Azure

We also need to choose the number of frequency bands, i.e. the resolution of the frequency axis. The number of frequency bands has a physical meaning – we cannot increase the number of frequency bands arbitrarily. For example, if we choose a small number of bands, say 10, when calculating the spectrogram, the spectral resolution will only be 10 units and the spectrogram will lose a lot of information (see the image at the left in Figure 3 – it has a very coarse representation of the original audio signal). On the other hand, if we choose too many bands, such as 1000 (the figure on the right), we will have a high-resolution image, but there will be many empty bands since we under-sample the signal within each band (not enough data samples per frequency band given the fixed bitrate of our audio) – this is shown by the empty black regions in the spectrogram. Choosing the number of bands is somewhat empirical too, and in this case, we choose 150 bands – a widely used number in many papers.