A sound meter reading taken with a machine running includes the machine and the surrounding sound. Subtracting the background means subtracting the underlying mean-square pressures under suitable measurement conditions. Subtracting the two decibel numbers gives a level difference, not the machine’s sound level.
The more difficult case comes when the total is only slightly above the background. A calculator can still produce a neat answer, but small changes in either input can move that answer substantially. This guide shows how to calculate a source estimate, recognise when it is unstable and design a more informative measurement. All numerical cases are original teaching examples, not site surveys or compliance assessments.
eduKateSG · The Gold Standard of Acoustics
Choose your reading route
Separate source from background · Chapters 1–2. Build the quantities before using a formula.
Calculate and stress the correction · Chapters 3–5. Follow original examples, assumptions and changed conditions.
Audit the measurement and continue · Chapters 6–8. Test transfer, answer common questions and follow connected reading.
Full chapter index · Practice and evidence checks · Return to Acoustics
Chapter index
Separate source from background · Chapters 1–2
Calculate and stress the correction · Chapters 3–5
Audit the measurement and continue · Chapters 6–8
CHAPTER 1 OF 8 · Separate source from background
1. Define the sound you want to estimate
Imagine checking whether a small ventilation unit has become noisier. With it operating, you measure the room. With it stopped, you measure the remaining sound: traffic, other equipment and the instrument’s own contribution. The second record is useful only if it reasonably represents what was present during the first.
Write Lₜ for the total level and Lᵦ for the background level. The desired quantity, Lₛ, is the source contribution at the microphone under the chosen operating condition. It is not automatically the machine’s sound power, a rating at another distance or what every person in the building experiences.
Keep the microphone position, instrument configuration, frequency weighting and analysis bandwidth consistent. Compare compatible averages over representative intervals. A maximum recorded during one event cannot simply be corrected using an unrelated average background.
There is a physical assumption too: source and background must combine with negligible time-averaged cross-correlation for simple energetic addition to apply. Two coherent loudspeakers can interfere. Their combined pressure may rise or fall with phase, so a source-on/source-off subtraction needs a different interpretation.
The Acoustics root develops the pressure, phase and correlation foundations. Here we focus on a narrower inverse problem: recovering one contribution from a combined measurement.
CHAPTER 2 OF 8 · Separate source from background
2. Undo the logarithm before subtracting
A sound pressure level can be written L = 10 log₁₀(q), where q is the mean-square pressure divided by the reference pressure squared. In air the conventional sound pressure reference is 20 micropascals. When measurements use the same reference and weighting, we can work with q without calculating pressure in pascals.
Under the additive model, qₜ = qₛ + qᵦ. Rearranging gives qₛ = qₜ − qᵦ. Convert each level to q, subtract, then return to decibels:
Lₛ = 10 log₁₀(10^(Lₜ/10) − 10^(Lᵦ/10)).
An equivalent form is Lₛ = Lₜ + 10 log₁₀(1 − 10^(−Δ/10)), where Δ = Lₜ − Lᵦ. This form shows that the correction depends on the separation between total and background. It also makes the domain visible: the expression requires a positive residual.
If Lₜ = Lᵦ exactly, the inferred residual under this ideal model is zero and its logarithm has no finite value. That does not establish that a real machine emitted no sound. The measurements may lack the resolution or comparability needed to separate it.
Brüel & Kjær’s Environmental Noise booklet, page 13 presents energetic background correction and cautions against correction when total and background are too close. Its guidance is a useful introduction; a particular testing procedure may impose its own acceptance rules.
CHAPTER 3 OF 8 · Calculate and stress the correction
3. Work a source estimate and check it forwards
Suppose matched measurements give 68.0 dB total and 62.0 dB background. Keep extra digits during calculation and round only the report.
| Quantity | Linear relative mean square | Meaning |
|---|---|---|
| Total, 68.0 dB | 6,309,573 | Source plus background |
| Background, 62.0 dB | 1,584,893 | Remaining sound estimate |
| Difference | 4,724,680 | Estimated source contribution |
| Return to decibels | 66.74 dB | Source level under the model |
The source estimate is about 66.7 dB. The arithmetic difference 68 − 62 = 6 dB describes the separation between the two records. It is not the answer to the source-level question.
Now check forwards. Convert 66.74 dB back to its linear value and add the background value. You should recover approximately 6,309,573 and therefore 68.0 dB. This check catches sign errors and confusion between factors of 10 and 20.
A second check is qualitative. A positive background contributes part of the total, so the estimated source level must lie below the total. It need not lie above the background: a sufficiently faint source can be masked by a stronger background.
A third check concerns interpretation. Even perfect arithmetic cannot show that the 62.0 dB background remained unchanged while the unit was running. That claim requires observations, repeat measurements or a justified experimental design.
CHAPTER 4 OF 8 · Calculate and stress the correction
4. See why a small separation magnifies uncertainty
Now use 63.0 dB total and 62.0 dB background. Their linear values are approximately 1,995,262 and 1,584,893. The residual is only 410,369, corresponding to 56.13 dB.
That value is mathematically consistent with the nominal inputs. Its stability is a separate question. To examine sensitivity, suppose each input might be displaced by 0.5 dB. This is an illustrative bounded-input exercise, not a statistical confidence interval.
| Input case | Total | Background | Estimated source |
|---|---|---|---|
| Nominal | 63.0 dB | 62.0 dB | 56.13 dB |
| Larger positive residual | 63.5 dB | 61.5 dB | 59.17 dB |
| Boundary residual | 62.5 dB | 62.5 dB | No finite logarithmic estimate |
For readers comfortable with calculus, the local sensitivities are ∂Lₛ/∂Lₜ = qₜ/(qₜ − qᵦ) and ∂Lₛ/∂Lᵦ = −qᵦ/(qₜ − qᵦ). At a 1 dB separation they are about +4.86 and −3.86. Each denominator contains the small residual.
With independent standard uncertainties of 0.2 dB on both inputs, first-order propagation gives about 1.24 dB standard uncertainty for this nominal source estimate. Near the boundary, the linear approximation can itself become poor and the output distribution asymmetric. Correlated errors also change the propagation; a common calibration offset is not the same as unrelated environmental fluctuations.
The practical lesson is to improve the measurement before adding decimal places. A stable subtraction needs information about both the source-on and source-off situations.
CHAPTER 5 OF 8 · Calculate and stress the correction
5. Challenge the assumption that the background stayed the same
Suppose the background during operation is actually 60 dB and the source contributes a linear value of 1,000,000, also 60 dB. Their combined level is about 63.01 dB.
Later, a passing vehicle raises the source-off background record to 61 dB. Subtracting that record from the earlier total yields a source estimate of about 58.70 dB. The source has apparently become quieter because the wrong background was used.
Reverse the mismatch and the bias reverses. A source-off record taken during an unusually quiet interval can make the source contribution appear too large. Background subtraction is an estimate of a counterfactual: what would the microphone have recorded at that time without the target source?
Useful improvements include bracketing an operating interval with source-off measurements, keeping a time history, recording operating states and identifying unrelated events. None guarantees that the background was identical. They make the assumption inspectable.
Frequency-band analysis may reveal a source that is obscured in a broadband summary. For example, a machine’s narrowband tone may be easier to distinguish from low-frequency traffic in the appropriate band. The microphone and bandwidth must still be suitable, and band results cannot be combined by ordinary decibel addition.
Do not use this method to explain active noise cancellation as removal of an independent energy contribution. Cancellation involves a deliberately correlated sound field. The cross-term that this simple model neglects is central to that mechanism.
CHAPTER 6 OF 8 · Audit the measurement and continue
6. Practise a changed-background investigation
Case A: a clear separation
A hypothetical source-on reading is 72 dB and a compatible background is 52 dB. Estimate the correction before calculating it.
The separation is 20 dB, so the background is one hundredth of the total linear mean square. The correction is 10 log₁₀(0.99), approximately −0.044 dB. The estimated source is 71.96 dB. At this separation the correction is small, although calibration, representativeness and other uncertainties remain.
Case B: a negative residual
An operating record gives 61 dB and a later source-off record gives 62 dB. Should the report say that the machine produced −1 dB?
No. A negative difference between the underlying mean squares cannot be assigned a real logarithmic source level in this model. Investigate background change, sampling variability, configuration and operating conditions. Report that the contribution was not resolved by this comparison; do not replace the negative residual with zero and imply a proven absence of sound.
Case C: the same total, a different background
Keep the 68 dB total but change the compatible background from 62 to 65 dB. The source estimate becomes about 64.98 dB. A larger fraction of the same total is now attributed to the background.
For a practical evidence record, preserve the raw records, interval lengths, microphone position, weighting, calibration checks, source state, environmental notes and the correction method. Include how close the records were and how uncertainty was assessed. A corrected number without these details is difficult to review.
Use supplied or simulated data for student exercises. There is no need to generate loud sounds to investigate logarithms or measurement instability.
CHAPTER 7 OF 8 · Audit the measurement and continue
7. Questions that prevent misleading reports
Is a low corrected level evidence of a quiet machine?
It may support that conclusion under an adequate measurement design. Near the background, a low estimate may instead mean the source contribution was poorly resolved. State both the estimate and its practical limitations.
Can I subtract the background in dBA?
The same energetic logic can apply to compatible A-weighted mean-square measurements. Preserve the weighting and time basis. A-weighted and unweighted levels are not interchangeable inputs.
Does a 3 dB difference guarantee a valid correction?
No. Separation is one diagnostic. The applicable measurement procedure, uncertainty, background stability and correlation assumptions still matter. Exact thresholds should come from the procedure being used.
Can a corrected sound level be below 0 dB?
A level below the reference is physically possible. That differs from a negative linear residual, which makes this particular subtraction undefined. The logarithmic reference and the sign of the residual are different issues.
Does measuring for longer always solve the problem?
Longer records can reduce some random variability. They can also span changes in traffic, weather or machinery. Choose intervals that represent the question and retain the time history.
CHAPTER 8 OF 8 · Audit the measurement and continue
8. Continue from an estimate to an acoustic explanation
Build the physical foundations
- The Gold Standard of Acoustics connects sound pressure, correlation, measurement and uncertainty.
- Sound, Decibels and Hearing Exposure develops logarithmic quantities and time averages; verify its measurement question before applying an exposure interpretation.
Keep neighbouring measurements distinct
- T20, T30 and What a Decay-Time Estimate Actually Measures explains a different use of levels: estimating decay from a selected part of a time record.
- Aperture Photometry and the Sky Counts Under a Star offers a useful comparison: background correction requires matching the contribution being removed to the measurement being interpreted.
Check the measurement model
- Brüel & Kjær: Environmental Noise, especially pages 11–13, explains energetic addition and subtraction.
- UNSW: Decibels and acoustic quantities connects pressure ratios and level references.
The transferable skill is to ask which physical quantity was added before trying to undo the addition. Follow The Gold Standard of Education: Learning That Transfers into Life and the How X Works hub for further routes through measurements, models and evidence.