What is Speech Privacy?
Speech Privacy measures whether conversations in one space can be overheard in an adjacent space. It combines two factors: how well the partition blocks sound (STC) and how much background noise masks intruding speech (NC).
Privacy Index (PI): PI below 80 = poor privacy (speech understood). PI 80-95 = normal privacy (speech audible but not intelligible). PI above 95 = confidential privacy (speech inaudible). Sound masking can improve privacy without changing walls.
PI = (1 − AI) × 100
That definition is ASTM E1130's. Everything below it is how this page estimates AI, and it is a broadband simplification rather than the band-by-band calculation the standard specifies:
received = SL − STC − 20 log₁₀(d ÷ 3 ft)
SNR = received − background
AI = (SNR + 12) ÷ 30, clamped to 0…1
SL is the vocal effort at 3 ft; d is the source-to-listener distance. The 30 dB span and the 12 dB offset are this tool's assumptions and have no published basis. Read the estimate as a screening figure, not a measurement — and note that PI hits its 100 ceiling for almost any real partition, which is why the Isolation Score is reported above it.
Isolation Score:
score = STC + background
This one is ours, and it is deliberately simple: no ceiling, one point per decibel, so it keeps separating options long after Privacy Index has pinned at 100. Use it to compare two designs on this page. Do not read it as a rating.
It is shaped like ASTM E2638's Speech Privacy Class, which is
SPC = LD + Lb — but the resemblance stops at the shape. E2638 takes both terms from measurements in the finished room, each averaged band by band across 160–5000 Hz. We have a laboratory partition rating and an A-weighted background level, which are different quantities in both terms. The number here is not an SPC and E2638's categories do not apply to it.
🎯 Privacy Index, and where it comes from
Privacy Index is not an independent measurement. It is a renormalisation of the Articulation Index: PI = (1 − AI) × 100, defined in an appendix to ASTM E1130. AI runs from 0, where speech carries no intelligibility, to 1, where every word is understood, so PI runs the other way — 100 is perfect privacy.
The Articulation Index itself is much older, growing out of work by Fletcher, Munson, French and Steinberg at Bell Labs in the 1930s and 1940s, and standardised as ANSI S3.5 in 1969. The Speech Intelligibility Index replaced it in the 1986 and 1997 revisions of that standard, adding hearing threshold and a modulation transfer function so that reverberation is accounted for. E1130 kept AI.
PI is not a percentage of words. The scale is not linear, and PI 80 does not mean 80 per cent of speech is unintelligible. Above PI 95 it saturates badly: Gover and Bradley's comparison of E1130 with the closed-room method found that sound insulation or background noise can vary by roughly 12 dB while PI moves only a few points.
📊 What the bands mean
The AI bands published with E1130, and the Privacy Index equivalents that follow from them:
| AI | PI | Description |
|---|---|---|
| 0.00 – 0.05 | 95 – 100 | Confidential speech privacy |
| 0.05 – 0.20 | 80 – 95 | Normal speech privacy |
| 0.20 – 1.00 | below 80 | Minimal or no speech privacy |
The 80 and 95 thresholds on the scale bar are those values. The further subdivision at PI 60, labelled “marginal”, is this tool's own split of the bottom band and does not come from the standard.
For reference, the five categories ASTM E2638 attaches to a measured Speech Privacy Class, with the Privacy Index each corresponds to. Both columns are Table 1 of Gover and Bradley; the PI column comes from the curve fit in their Figure 1. These thresholds belong to an E2638 measurement and cannot be applied to the Isolation Score above — that score adds an A-weighted background level where E2638 uses a band average, and an A-weighted total runs several decibels higher than the mean of the bands it is summed from:
| SPC | E2638 category | Equivalent PI |
|---|---|---|
| 90 | Very high speech security | 100 |
| 85 | High speech security | 100 |
| 80 | Standard speech security | 100 |
| 75 | Standard speech privacy | 98 |
| 70 | Minimal speech privacy | 93 |
The right-hand column is the whole argument in one place. Three different categories — a twenty-decibel span, from a room where speech is occasionally audible to one where it is not detectable — all report PI 100. Going the other way, the same table gives E1130's own two categories as PI 80 “normal” = SPC 64 and PI 95 “confidential” = SPC 71, so E1130's confidential threshold lands one point above E2638's lowest named category.
⚠️ The limitation that matters most
E1130 is a test method for open plan spaces — workstations separated by partial-height screens. It explicitly does not apply to closed rooms: private offices, conference rooms, exam rooms. The metric for a closed room is Speech Privacy Class, from ASTM E2638, computed as SPC = the level difference between source and receiving positions plus the background level in the receiving room, both arithmetically averaged over the 160 to 5000 Hz one-third octave bands. E2638 is a measurement procedure carried out in a finished space, not a design prediction.
This calculator takes a partition STC as its input, which is a closed-room situation. That is why it reports an Isolation Score as well, and why the Privacy Index below it will usually read 100.
Privacy Index cannot compare partitions. PI is capped at 100 by its own definition, and a closed room reaches the cap easily. At this tool's defaults — STC 45, 40 dBA background, 10 ft, a normal voice — every partition from STC 22 to STC 70 returns PI 100 and the same “confidential” label. Only STC 20 and 21 move it. A number that reads the same for a bathroom door and a double-stud wall is not a screening estimate; it is a constant.
This is not a quirk of our simplification. It is the documented behaviour of the metric, and it is why ASTM published a second one. Gover and Bradley, comparing the two across 100 simulated cases, found that “the sound insulation or noise can vary by ~12 dB, and yet PI varies by only a few percent”, that “all conditions higher than SPC 75 correspond to conditions for which PI > 98%”, and that within that range the fraction of listeners who could hear speech at all ran from 100 per cent down to zero. Their conclusion is the reason the box above exists: SPC “offers practicality in that a difference in, for example, 5 dB of sound insulation will correspond to a difference of 5 in SPC, whereas the corresponding difference in PI depends on the absolute value, and could be 0–2% for conditions of high privacy.”
So read the Isolation Score when you are comparing walls or deciding whether masking or construction is the better decibel, because it moves one point per decibel over the whole range. Read PI only when it is below 100, where it still has resolution, or when the space is genuinely open plan. If a room has to be demonstrated private, neither will do — that is an E2638 measurement in the finished space.
Why the score is not called a Speech Privacy Class. It has the right shape — E2638's SPC is the source-to-receiver level difference plus the background level — but the background term here is not the one E2638 uses. E2638's Lb is the arithmetic mean of the one-third octave band levels from 160 to 5000 Hz; the field on this page takes an A-weighted broadband level, which is an energy sum across those same bands and therefore sits well above their mean. For a background shaped like NC-35 the bands average about 34 dB while the A-weighted total is about 43 dBA. Adding the larger figure inflated the result by roughly nine points — one to two whole categories — and it did so in the direction that tells you a room is more private than it is. Recovering the band average from a single dBA number requires assuming a spectrum the user has not given us, and that assumption would be doing all the work, so the categories were removed rather than patched.
Two further reasons the number is an estimate rather than a rating. AI in E1130 is computed band by band with speech-importance weights; this page collapses that to a single broadband signal-to-noise ratio. And the path loss here is a partition STC plus inverse-square spreading from a 3 ft reference, whereas real room-to-room noise reduction also depends on the area of the partition and on how absorptive the receiving room is. A hard, bare receiving room is worse than the same wall in a carpeted one, and that is not modelled.
🛠️ The two levers, and which one is cheaper
Privacy is a ratio, not a level. Every point of it comes either from reducing what arrives — a better partition, sealed penetrations, attention to the ceiling plenum and shared ducts — or from raising what masks it. Because the two enter the calculation symmetrically, a decibel of background noise is worth the same as a decibel of extra transmission loss, and background noise is usually the cheaper decibel by a wide margin.
That is the whole argument for sound masking. It is also why so many otherwise good partitions disappoint: the wall was built to STC 50 and the room it protects sits at 30 dBA, so speech that gets through has nothing to hide in. Note that masking has its own ceiling — push it high enough to guarantee privacy and the occupants complain about the noise instead.
For a sense of scale on what background noise is acceptable where listening matters, ANSI/ASA S12.60/Part 1-2010 caps background noise in an unoccupied furnished classroom at 35 dBA. That is a limit for a room where speech must be understood — the opposite problem — and it is a useful reminder that the same decibel does opposite work depending on which side of the wall you are standing.
❓ Common questions
How is speech privacy defined?
By whether speech from one space can be understood in another, measured rather than judged. ASTM E1130 defines the Articulation Index for open plan spaces, running from 0 for unintelligible to 1 for fully intelligible, and defines Privacy Index as PI = (1 − AI) × 100 so that higher means more private. For closed rooms the metric is Speech Privacy Class under ASTM E2638, which adds the source-to-receiver level difference to the background level in the receiving room across the 160 to 5000 Hz third-octave bands. The two are different methods for different situations and their numbers are not interchangeable.
How do you calculate speech privacy potential?
You need three things: how loudly people speak, how much the construction between them reduces that speech, and how much background noise is present where the listener is. Privacy comes from the ratio of the last two to the first, not from any one of them alone. A calculator can estimate this well enough to compare options, but the standardised methods are measurements taken in the finished space, and a room that has to be demonstrated private needs an ASTM E2638 test rather than a prediction.
What does acoustic privacy mean?
It means a conversation cannot be understood outside the space it happens in — which is not the same as it being inaudible. Normal privacy, in the ASTM E1130 bands, corresponds to a Privacy Index of 80 to 95, where speech may well be audible but only the occasional word is picked out with effort. Confidential privacy is PI 95 or above, where nothing intelligible gets through. Achieving either depends as much on the background noise in the listening space as on the wall.
How to mask sound for privacy?
By raising the background noise in the listening space with a broadband, steady, spectrally shaped sound distributed evenly through the ceiling — that is what a sound masking system does, as opposed to a white noise machine on a desk, which is neither even nor shaped. Evenness matters more than level: a masking sound you can localise or that changes as you walk becomes the distraction. Masking works because privacy depends on the ratio of intruding speech to background noise, so a decibel added to the background counts the same as a decibel of extra transmission loss and usually costs far less.
Do white noise machines work for office privacy?
Partially, and only nearby. A single machine raises the background noise close to it, which does improve the speech-to-noise ratio at that one position, but the level falls off with distance and the spectrum is not shaped for speech masking, so coverage is uneven and it can be more annoying than the speech it is covering. Engineered masking systems distribute a shaped sound evenly across a whole area for that reason. For a closed office, sealing the obvious leaks — the door undercut, the ceiling plenum above the partition, back-to-back outlets — usually buys more privacy per pound than any noise source.
The Privacy Index definition PI = (1 − AI) × 100, the AI and PI bands, and the scope limitation of ASTM E1130 to open plan spaces are taken from Cambridge Sound Management's Speech Privacy Standards technical note and from B.N. Gover and J.S. Bradley, ASTM metrics for rating speech privacy of closed rooms and open plan spaces, Canadian Acoustics 39(3), 2011, which also supplies the ASTM E2638 Speech Privacy Class definition and the observation that PI saturates above 95. Background-noise limits for learning spaces are from ANSI/ASA S12.60/Part 1-2010 (R2020). ASTM E1130, ASTM E2638 and ANSI S3.5 are named to identify what governs each metric; all are paywalled and we have not read them.