Noise is one of those subjects in audio that always comes up. There is a lot of good information out there, and quite a bit of misinformation as well. A microphone may advertise 7 dBA self-noise. A preamp may claim an EIN of –129 dBu. One converter may offer 120 dB of dynamic range while another records in 32-bit float. Then we take all of this supposedly quiet equipment into a room, turn up the monitors, and hear hiss.
So where is the noise actually coming from? To answer that, we need to start with the physics.
Noise starts with temperature
Thermal noise comes from random microscopic motion in matter. In a gas, molecules are in constant motion and collision. In a conductor, charge carriers fluctuate randomly. Brownian motion and Johnson–Nyquist noise are two manifestations of the same broader idea: at any temperature above absolute zero, a dissipative physical system has thermal fluctuations.
At a comfortable room temperature of about 20 °C, the absolute temperature is roughly 293 K. That thermal energy is always present. Cooling a system reduces thermal noise, but ordinary audio equipment does not get to opt out of thermodynamics.
For an electrical resistance R, the RMS Johnson–Nyquist noise voltage over a bandwidth Δf is
Here, kB is Boltzmann’s constant, T is absolute temperature, R is resistance, and Δf is measurement bandwidth. You do not need to plug numbers into this every time you record. The important lesson is what the equation says: more resistance, more bandwidth, or more temperature means more integrated thermal noise.
A theoretically perfect resistor at room temperature still makes noise.
A microphone capsule has thermal noise too
A condenser microphone has an acoustic version of the same problem. The diaphragm, the thin layers of air around it, the backplate holes, vents, damping structures and other losses can be represented by mechanical and acoustic impedances. Wherever there is a real — dissipative — part of that impedance, thermal agitation produces a corresponding random force or pressure.
If the relevant loss is expressed as an equivalent acoustic resistance Ra referred to pressure at the diaphragm, the pressure-noise relationship has the same form:
This is an important point: some of the microphone’s self-noise exists before the electronics ever see the signal. Making the impedance converter quieter cannot remove thermal noise already generated by the capsule’s acoustic losses.
Capsule geometry, damping, diaphragm area and sensitivity therefore matter enormously. A larger diaphragm does not somehow eliminate Brownian noise. What it can do is improve the ratio between the useful electrical output and the underlying noise mechanisms.
How noise adds
The next question is what happens when several noise sources are present at the same time.
If two copies of the same coherent signal have identical phase and amplitude, the amplitudes add directly. Double the amplitude and the level rises by 6.02 dB.
Independent thermal-noise sources are different. Their instantaneous waveforms do not track one another, so their powers add rather than their amplitudes.
For two independent noise sources with the same RMS level, the result is √2 times one source, which is only 3.01 dB higher. This one distinction explains a great deal about practical noise budgets.
If one noise source is substantially larger than the others, it dominates the total. Making an already-small contributor even smaller produces very little improvement.
The ear does not hear noise as one big number
Our hearing also complicates the story. The cochlea does not analyze the entire audio spectrum as a single broadband level. It behaves approximately like a bank of overlapping auditory filters.
The classic Bark model describes about 24 critical bands. Modern auditory-filter models refine the details, particularly at low frequencies, but the central idea remains useful: sounds that occupy the same auditory band mask one another much more strongly than sounds that are well separated in frequency.
This explains a familiar experience. You can sometimes hear a bird above a rushing river or pick a conversation out of a crowded room even when the broadband background level is high. The signal has spectral and temporal structure that does not completely overlap the masker.
A pure tone concentrates most of its energy in a narrow frequency region. White noise spreads its energy across the spectrum. Two sounds can therefore have the same overall SPL and very different perceptual effects.
There is no honest universal rule saying that a signal exactly 10, 15 or 20 dB below another sound will always disappear. Masking depends on frequency separation, bandwidth, level, duration and the character of both sounds.
What does dBA actually mean?
Microphone self-noise is commonly specified in dBA. The “A” means the measurement has been passed through the standard A-weighting curve, which reduces the contribution of frequencies to which human hearing is relatively insensitive — especially very low frequencies.
That makes A-weighting useful, but it also means a dBA number is not simply the total physical noise energy coming from a microphone.
As a useful reference, if perfectly white noise is flat in power per hertz from 20 Hz to 20 kHz, applying the standard A-weighting curve reduces the integrated level by about 2.0 dB. Real microphone noise is not perfectly white, so the difference between weighted and unweighted measurements depends on the actual spectrum.
Why microphone S/N numbers look so good
Microphone signal-to-noise ratio is normally referenced to a sound pressure of 1 pascal. Because the SPL reference pressure is 20 µPa, 1 Pa corresponds to approximately 94 dB SPL:
A microphone with 14 dBA equivalent self-noise therefore has a published A-weighted signal-to-noise ratio of roughly 80 dB when the reference signal is 94 dB SPL.
But many things we record are nowhere near 94 dB SPL at the microphone. If a quiet source is only about 44 dB above the same reference and the microphone contributes the equivalent of 14 dBA noise, the available S/N is roughly 30 dB — assuming comparable weighting and bandwidth.
The microphone did not get noisier. The signal got quieter.
This is why self-noise matters far more for quiet ambience, distant orchestral pickup, nature recording and other low-level material than it does when recording a snare drum or a guitar cabinet.
The condenser microphone noise budget
Inside a condenser microphone, the two major contributors are usually the capsule and the impedance converter. The capsule creates a tiny, very high-impedance electrical signal. The impedance converter buffers that source so it can drive the rest of the microphone and a cable.
A well-designed impedance converter should contribute substantially less noise than the capsule itself. Because independent noise sources add in power, once the electronics are comfortably below the capsule, additional improvement gives diminishing returns.
If one contributor is 10 dB below the dominant noise source, it raises the total by only about 0.4 dB. At 15 dB below, the increase is only about 0.14 dB.
I think of this as engineering invisibility: once a stage is far enough below the dominant limitation, making it still quieter may look excellent on a data sheet while changing almost nothing we hear.
What about the microphone preamp?
Microphone preamp noise is commonly specified as Equivalent Input Noise, or EIN. Good modern preamps often land around –127 to –130 dBu under favorable measurement conditions.
For perspective, the Johnson noise of a 150 Ω source at 293 K integrated from 20 Hz to 20 kHz is about 0.22 µV RMS, or roughly –131 dBu. Once a preamp approaches this region, its designers are fighting the thermal noise of the source resistance itself.
There is an important caveat when comparing specifications: dBA microphone self-noise and dBu preamp EIN are not directly comparable numbers. To compare them properly, the microphone’s acoustic self-noise must be translated through its sensitivity into an equivalent output voltage.
For most modern condenser microphones connected to a competent low-noise preamp, the preamp is not the biggest contributor. It amplifies the microphone signal, amplifies the microphone’s own noise, and then adds a smaller amount of its own noise.
And the A/D converter?
Digital resolution often gets mixed into the noise discussion. For an ideal N-bit converter driven by a full-scale sine wave, the familiar theoretical quantization S/N is
That works out to about 98 dB for 16 bits and about 146 dB for 24 bits. Real 24-bit converters do not deliver 146 dB of usable analog dynamic range because the analog circuitry, voltage references, resistors and input stages all live in the same thermal world we have been discussing.
Even so, modern converters routinely provide more than 100 dB of real dynamic range. Once the microphone signal has enough analog gain ahead of the converter, A/D noise is usually well below the microphone and preamp noise.
Also remember that 0 dBFS is a digital maximum, not one universal analog voltage. The analog voltage corresponding to digital full scale is defined by the particular recorder or interface.
What 32-bit float actually changes
This subject creates a surprising amount of confusion.
32-bit floating-point recording and the A/D converter are not the same thing. Thirty-two-bit float describes the numerical representation used to store the audio. Many modern field recorders pair that format with two or more A/D paths operating at different analog gains.
One path handles quiet material with greater gain. Another preserves headroom for loud material. The recorder combines or selects those paths digitally, then stores the result in a 32-bit floating-point file with enormous numerical headroom.
This largely solves a level-setting problem. You can capture a very wide range of input levels without riding the gain knob in the traditional way.
It does not defeat physics. The microphone still has self-noise. The analog front end still has noise. The capsule, impedance converter and input circuitry still have maximum levels. If one of those analog stages overloads before the digital conversion is complete, no numerical file format can reconstruct the missing waveform.
Do cables add noise?
Technically, every conductor with resistance has thermal noise. In a normal balanced microphone cable, however, that contribution is tiny. Cable problems in real systems are much more often interference, shielding, connections or physical failure.
A cable near a lighting dimmer or another strong electromagnetic source can pick up unwanted energy. RF from transmitters or phones can enter a poorly protected system and be demodulated by nonlinear junctions. A damaged shield can destroy the common-mode rejection that balanced connections depend on. Connectors can oxidize, loosen or become intermittent.
Those are real problems.
Whether the copper contains a few hundred parts per million of oxygen or less than ten is not a meaningful audio-frequency performance issue in an ordinary microphone cable. Oxygen-free copper has legitimate manufacturing uses; it is not a magic noise-reduction technology.
For microphone cable I care about conductor resistance, shield quality and coverage, twist geometry, connector quality, flexibility and strain relief. I use gold-plated XLR contacts because gold resists corrosion — not because gold somehow improves the music.
So why do I still hear hiss?
This brings us back to the original question.
Sometimes the microphone is the limiting factor. Sometimes the preamp matters. Sometimes there really is interference. But remarkably often, what we call “noise” in a recording is simply the environment.
HVAC systems move air. Buildings vibrate. Traffic produces low-frequency energy. Lighting systems buzz. People move. Air travels through ducts. Wind interacts with microphones. Electronics radiate electromagnetic energy.
Move a microphone farther away from an orchestra and the orchestra becomes quieter at the microphone while much of the room noise does not. The equipment has not changed. The acoustic signal-to-noise ratio has.
That is why a microphone with spectacular self-noise specifications cannot make a noisy room quiet. It is also why moving the microphone can sometimes improve the recording more than buying a preamp with another decibel of EIN performance.
Follow the largest noise source
When I troubleshoot noise, I work from the largest contributor downward.
Start with the acoustic environment. Then look at the microphone capsule and its sensitivity. Then the impedance converter. Then the preamp. Then the converter. Check cables and interference when the character of the noise suggests a problem rather than ordinary broadband hiss.
Once one source dominates the system, spending enormous effort improving everything underneath it accomplishes very little.
That is the real lesson behind the equations. We cannot eliminate thermal noise, but once we understand where the noise comes from, how independent sources combine, and how our hearing responds to them, we can stop chasing the wrong numbers and concentrate on the parts of the recording chain that actually matter.
References & further reading
- NIST — Noise, Temperature, and the New SIJohnson noise and thermal motion.
- Analytical-numerical model of a condenser microphoneMechanical-thermal and acoustic-resistance noise relationships.
- Neumann — What is Self Noise?94 dB SPL / 1 Pa reference and microphone S/N.
- Development and Current Status of the Cambridge Loudness ModelsAuditory filters, critical bands and ERB refinements.
- Zoom F6Dual A/D conversion and 32-bit floating-point recording.