Linear and Logarithmic Scaling

Prev Next

One issue that you run into repeatedly in acoustical analysis is that human perception is logarithmic in nature and covers a relatively huge range of values. Everyone’s hearing is a little different but in general, the difference between the threshold of hearing and the threshold of pain – the quietest sounds we can hear and the loudest sounds we can stand – is somewhere around 120 decibels (dB).

That works out to six orders of decimal Magnitude (the difference between one and one million, e.g.).

In terms of frequency, the audible Spectrum for humans is typically defined as 20 Hz to 20 kHz, a range of four logarithmic “decades”. Admittedly, many or perhaps most of us are unable to hear across that entire range but it might be safe to say that most people can hear across a range of at least three decades, e.g., from 80 Hz to 8 kHz, which is still a pretty wide range of numbers.

The thing is, we do not hear differences between all those numbers equally in either case. To our senses, the difference between one and two is not the same as the difference between two and three, as it would be if we perceived the world linearly. To us, the difference between one and two sounds (or looks, or feels) more like the difference between two and four, or four and eight, or eight and sixteen… (you get the idea).

Charting audio and acoustic data on logarithmic Amplitude (magnitude) or frequency scales does two useful things for us then; it helps to make the wide ranges of values that our hearing encompasses more manageable and it results in a presentation of the data that is often more meaningful in terms of human perception. None of this is to say that linear scales don’t have their uses, but for most of the things we do in Smaart, logarithmic scales and units (decades, octaves and decibels), tend to do a better job of showing us what we want to see in a way that makes intuitive sense.

Linear and Logarithmic Frequency Scales

When we talk about Linear and Logarithmic frequency scales (not to be confused with fractional octave banding) we are really just talking about how frequencies are plotted on charts and graphs. On a linear frequency scale, let’s say every 100 Hertz (you can pick any number), occupies the same amount of space on the chart as every other. On an octave scale, each octave is the same width as every other, even though the linear frequency range for each band doubles as you ascend in frequency (125 Hz, 250 Hz, 500 Hz, 1 kHz, 2 kHz, 4 kHz…). On a logarithmic decade scale, each power of 10 Hz, (10, 100, 1000,10,000) is the same width as every other. Logarithmic scales work the same way for any base, but the bases we use for log scales in Smaart are two and ten (octaves and decades).

Linear vs Logarithmic frequency scaling. Two views of the same comb filter on a linear and log scaled magnitude graph.

We most often look at frequency on octave or decade scales because these correlates better with our own logarithmic perceptions of sound. However, linear scales are very useful for some things as well, and sometimes correlate better with the underlying physics of sound and acoustics. Charting frequency on a Linear Scale can make comb filters and harmonic distortion products stand out more clearly since the lobes or peaks are linearly spaced. Another example might be the Phase shift associated with a fixed delay, which becomes a straight-line slope on a linear frequency scale.

Note that when you look at FFT data from acoustical measurements or other noisy signals on a log frequency scale, the trace gets fuzzier-looking at higher frequencies. That doesn’t necessarily mean there is more noise in the HF. It is a natural consequence of packing more and more linearly-space FFT points into a smaller and smaller amount of chart space. That is one of the reasons for the MTW Transfer Function option, as noted earlier. Smoothing also helps to reduce visual noise in the HF in transfer function measurements, as does fractional octave banding for spectrum measurements.

MTW vs 16K FFT transfer function on a logarithmic frequency scale. The MTW uses larger time constants at low frequencies to improve LF resolution while smaller time constants at higher frequencies reduce visual “noise” due to excess resolution.

Linear Amplitude

Linear amplitude, as the name might imply, is amplitude displayed on a linear scale, e.g., volts or digital integer-based amplitude units. In Smaart, the only places that you ever see linear amplitude are linear time-Domain charts, where amplitude is displayed as a percentage of normalized Full Scale. That is to say that the largest positive and negative numbers obtainable from a signed integer of given number of bits (e.g., 16 or 24 bits per sample) are scaled to a range between 1 and -1 (inclusive), with fractional values in between expressed as percentages.

Since you can’t take the log of a negative number, the only way to see relative polarity in an Impulse Response is to use a linear amplitude scale. Also, some people prefer the linear amplitude scale for identifying discrete reflections in an impulse response, and it can be useful for looking at other types of signals as well. A linear amplitude scale tends not be very useful for looking at reverberant decay or for identifying peak structures in the LF range of an impulse response where the length of a waveforms period is spread out of time so much that a clear impulse is not easily discernable.

Linear vs Logarithmic amplitude scaling. The impulse response of a bandpass filter is shown on a linear (percentage of normalized full scale) versus logarithmic (Decibel) amplitude scale. Notice that only the first two oscillations in the IR are easily discernable on the linear (Lin) view, whereas the Log view clearly shows the first six corresponding lobes.

Decibels (dB)

The decibel is a logarithmic ratio commonly used to express amplitudes, voltages, sound pressure, gain and Attenuation and no doubt other things as well. The word literally means one tenth of a Bel. The Bel is named for Alexander Graham Bell, inventor of the telephone (or one of the inventors anyway). Why they called it a “Bel” instead of a “Bell” is a question that someone else would have to answer, but that explains why the abbreviation for decibels is written as dB (with a capital B). Although no one seems to use Bels for much of anything – most people would just say 10 dB instead – the formulas for converting to and from decibels may seem less arbitrary if you consider that one Bel represents the logarithm of a power ratio of 10:1 and a decibel is 1/10th of that.

With that thought in mind:

Because decibels express a ratio, decibel values must be referenced to something. If no reference is explicitly given, as in the equations on the previous page, the reference is assumed to be one, however, it could potentially be any number. To reference dB to a number other than one, you simply divide the value that you want to convert to dB by the reference value before taking the log.

In that case:

One common example of this in audio applications is dBu, which references 0 dB to 0.775 Volts. Another might be the AES convention for dB FS (dB Full Scale) which essentially references 0 dB to the square root of 0.5 (0.7071), so that a full-scale, peak-to-peak sinewave has an RMS value of 0 dB instead of −3.01 dB. Note that Smaart does not use the AES convention for Full Scale.