Understanding Sound Waves and Analysis Methods

Sound as Pressure Waves

Sounds are pressure waves that propagate through a medium. They are generated by vibrating sources and encode valuable information about the characteristics of the vibrations.

At Resonyx, we collect these sound waves using high-sensitivity microphones and convert them into digital signals. Depending on the application, we work with sampling frequencies between 20-40 kHz, meaning we capture between 20,000 - 40,000 pressure measurements per second. These digital signals can be stored, replayed, or further analysed to extract critical information about the source.

Extracting Features from Sound

With thousands of data points per second, using raw sound directly in end-to-end analysis models is impractical. The first step in our analysis pipeline involves extracting relevant features from the signal, which are then analysed by specialised models. Our system contains a broad catalogue of feature extractors that operate in real time on our edge device. These extractors fall into two major categories: features obtained directly from the waveform and those derived from the sound spectrum (covered in the next blog post).

One fundamental characteristic of a waveform is its amplitude, which directly relates to  the total energy of the wave, which in turn translates into loudness. Loudness is computed within fixed-size windows, typically ranging from tenths of a second to a few seconds, depending on the application. This process drastically reduces data volume—from tens of thousands of data points per second to just a few per second. However, because a significant amount of information is discarded, loudness alone is rarely sufficient for detailed analysis but provides an intuitive indicator of the general behavior of the sound.

In addition to loudness, our system extracts other time-domain features such as waveform envelope detection, which connects local maxima to form an amplitude representation.