Noise performance is important, but to fairly compare different products, you need to understand how noise specs are calculated.
Two primary signal‑processing phenomena lead to audio‑signal degradation: Distortion (which will be the subject of Part 3 of this series) and Noise, the subject of this instalment. Noise and distortion figures provided in manufacturer’s spec tables, therefore, are key quantities when comparing the performance of two competing audio signal‑processing products.
We can differentiate and define the two phenomena for the purposes of this treatment. Noise is any signal added to and independent of the intended audio by elements of the signal path. As such, noise is the signal present in the absence of an audio source, as is the case with a terminated or shorted input. Distortion, by contrast, is any signal component added by elements of the signal path in response to the intended audio. In the absence of an audio input, distortion is zero. In any but minute amounts, both phenomena reduce the clarity, presence and realism of audio presented to a listener.
In sufficient quantities, both can impact intelligibility, so they are important quantities not only in audio production and high‑quality reproduction environments but even in what we might think of as lo‑fi applications like, for example, transit‑system PAs, or mobile communication systems used by emergency first responders and commercial fleets.
At minimum, noise performance limits a system’s dynamic range — the range from the smallest to largest audio‑signal amplitude a system can represent. In particular, noise sets the lower limit of the dynamic range, establishing what engineers refer to as the noise floor. As well as reducing the intelligibility of speech — particularly soft or similar‑sounding phonemes — noise can mask the performance nuances of acoustic musical instruments. In recording and post‑production environments, dynamics processors can exacerbate the effects of noise allowed in at early stages of the signal chain, no matter if dynamics are managed in the analogue or digital domain.
Once added to a signal, noise is indistinguishable from intended audio by real‑time signal processors.
Once added to a signal, noise is indistinguishable from intended audio by real‑time signal processors. So, for example, gain stages amplify input noise as much as input signal. For this reason, the noise spec on signal‑processing blocks that provide large amounts of gain, such as a mic preamp or instrument channel, may be more concerning than on a signal processor that typically provides either unity gain or only several dB of ‘make‑up gain’ as with, for example, an analogue equaliser.
Audio‑equipment manufacturers typically specify the noise performance of their products in one or more of three ways: as an absolute Noise amplitude, as an SNR (signal‑to‑noise ratio), or as a Dynamic Range. The absolute noise amplitude is the most basic of these and is, in fact, necessary for manufacturers to measure and used to calculate the other two.
Though the broad subject of noise is sufficiently complex to engender textbook‑length deep dives, for the purpose of understanding audio‑equipment spec sheets, we need only scratch the surface of a few concepts.
Among them is the fact that the dominant noise sources in modern audio equipment tend to share a common spectral shape — that is the curve revealed by the way their energy distributes across a frequency range or spectrum. In audio gear, dominant noise sources tend to distribute their energy evenly across the audio‑frequency range. Referred to colloquially as white noise (being analogous to white light, which similarly distributes its energy evenly across the visual spectrum), these sources give rise to a mostly benign‑looking expression of noise‑power density in units of (fractional) W/Hz RMS. Here, the Hz refers to the measurement bandwidth. The wider the bandwidth, the more noise is detectable by the measurement.
As we saw in Part 1 of this series, terms and conditions apply and, here, a key condition is the measurement bandwidth, which spec sheets should identify. When considering competing products, check that the noise‑measurement bandwidths match to make direct comparisons possible. Often, you’ll find that the measurement bandwidth is 20Hz to 20kHz or 20Hz to 22kHz. Tests made with unmatched measurement bandwidths are not directly comparable, though one can calculate estimates with a certain amount of mathematical machination in the case where we assume the dominant sources in both products are white noise. (It’s also important to check that both measurements are taken using the same weighting scheme, a complication to which we’ll return in a couple of months.)
A slight complication arises, however, when we note that we rarely measure or consider signal power. In most audio work, signals manifest as voltage amplitudes or their digital representations. We want to express noise, either directly or in comparison, in terms compatible with how we describe signals because the signal domain sets the context for our evaluations.
Signal power is proportional to signal voltage squared. To normalise measures to signal voltage, we take the square root of the power spectral density, expressing the noise voltage density in units of (fractional) V/√Hz RMS. Don’t let the square root scare you off. It’s just keeping the dimensions straight as we move from describing signals in the power domain to the voltage domain. You’ll likely never have reason to calculate the square root of the noise‑measurement bandwidth as a purchaser and user of audio gear.
Now that we’re in the voltage domain for both signal and noise, we need only adjust our scale, because we don’t use voltmeters to measure signals in audio production or post‑production environments; we use meters calibrated in dB.
Though a measure in dB always suggests a ratiometric quantity, equipment makers can stipulate a reference amplitude and calculate the noise relative to that. So, in many spec sheets, you’ll see noise expressed in dBV (dB relative to 1V) or dBu (dB relative to 0.775V — a historically meaningful reference potential that shows up often). Recall that 20dB corresponds to an amplitude ratio of 10:1. So, for example, a noise spec of ‑80 (= ‑4 x 20) dBu corresponds to a noise amplitude of 10‑4 x 0.775V = 77.5μV. If you’re uncomfortable with maths, don’t worry: there won’t be a test and you’ll not need to convert between voltages and dB. Equipment makers will do that for you.
One common approach to noise specs that manufacturers use is to express their product’s noise performance, not in absolute terms as we’ve done thus far, but relative to the product’s Reference Level (in analogue equipment) or, the maximum signal amplitude (in digital devices). Let’s take the noise spec we just looked at and apply it to an analogue product with a reference level of +4dBu. The resulting Signal To Noise Ratio (SNR) would be the subtraction of -80dBu from +4dBu, giving 84dB.
If, instead, that noise figure was derived from a digital converter where 0dBFS equated to +24dBu, the resultant signal-to-noise ratio would be the subtraction of -80 from +24 giving an SNR of 104dB. Note that, by using the dB’s logarithmic scale, we accomplish the calculation of a ratio, which implies arithmetic division, with simple subtraction. Note, too, that the SNR expression does not include the reference voltage suffix, because that quantity cancels when taking the ratio.
The remaining expression of noise performance in common use on audio‑equipment spec sheets, Dynamic Range (DR), is only slightly more conceptually complicated than SNR. In digital equipment the terms are essentially synonymous, whereas in analogue equipment SNR is always less than DR by the amount of available headroom. Many digital and mixed‑signal equipment designs, however, mute their outputs in the absence of an input signal, which would result in a noise measurement that fails to reflect the equipment’s performance in normal use.
To make realistic noise measurements in such cases, test systems first establish the maximum input signal as they would for a standard SNR measurement but, instead of setting the test input to zero, the input is set 60dB lower than that for full scale. The measurement system must then apply a sharp notch filter to the output to eliminate all traces of the stimulus signal and measure the remaining noise.
Keep in mind, when comparing noise specs between competing gear, to make your evaluations in the context of your audio application. Applications such as recording spoken word, be it for a podcast or audio for video, acoustic instruments either in‑studio or live, or room mics in a concert hall for ambience and audience response are all likely going to demand greater amplification and, thus, depend on good noise performance in the first gain stages than, say, channels processing mics over hard‑hit drums or taking large output signals provided by electronic instruments.
This series is produced in association with Audio Precision, Inc.
Microphones are analogue devices and, like most analogue devices, they tend to produce increasing distortion levels as the signal nears overload. This makes it difficult to define the maximum signal level: is that when THD reaches 0.5%, or 1%, or 3%? Consequently, it is more reliable — and of more practical relevance in normal use — to measure SNR from a defined Reference Level, which is generally where the normal signal level resides. In an analogue console that might be +4dBu, for example, while in a tape recorder it might be 320nWb/m. In microphones the reference level is always taken as 1 Pascal which equates to 94dB SPL.
- Signal To Noise Ratio for microphones is defined as the difference between the noise floor and a reference acoustic level of 94dB SPL. Since most studio microphones are specified to handle at least 120dB SPL (for 0.5% THD), this means that specified microphone SNR is typically much lower than the microphone’s Dynamic Range.
- The Dynamic Range of a microphone is defined as the ratio between the loudest and quietest acoustic signals to which the microphone responds in a reasonably linear fashion — respectively, the noise floor and the point where THD reaches 0.5% (although some manufacturers use 1% to get a better number for marketing purposes!).