Public Address & STI / STIPA, Episode 29

Key Takeaways

What you'll walk away with

Four ideas to carry into the next specification conversation you have.

STI is a 0-1 score for speech clarity.

STI turns speech into a test of rhythm: how much of the rise and fall of loudness, at 14 modulation rates in 7 octave bands, survives the room and the noise. Full STI reads all 98 combinations; STIPA reads 14 of them in one 15 to 25 s reading, which repeats to about ±0.03. That is why the standard's bands are 0.04 wide, and why 0.52 and 0.54 are the same result.

≥0.45 at 90% of locations, ADS average ≥0.50.

NFPA 72 Annex D is a two-part statistical test: at least 90% of measurement locations in each acoustically distinguishable space (ADS) at STI 0.45 (CIS 0.65) and the space average at not less than 0.50 (CIS 0.70). Read the status with the numbers: NFPA's own Manual of Style puts the same boilerplate on every informational annex, that it is not part of the requirements of the document. The mandatory text in Chapter 18 says quantitative measurements shall not be required, and shall be permitted. So these two numbers are the industry's working criterion, and they bind a project when its specification adopts them.

STIPA is the on-site test method.

It is, in ordinary rooms. With distinct echoes, loudspeakers spaced far apart without delay lines, or more than about 3 s of reverberation through a band-limited system, use Full STI instead. On duration, separate the standard from the meters: IEC 60268-16 times one measurement at 10 to 15 s, and about 15 to 20 s in its STIPA annex, while field meters run 15 to 25 s. The 30 seconds still quoted everywhere was RASTI's, and RASTI was marked obsolete in 2011 and deleted in 2020. The standard also bans, in normative language, estimating a predicted STI by converting a %ALcons value.

The room is nearly three times the lever.

Placement does matter, and the spacing rule is trigonometry rather than a multiple of ceiling height: spacing = (mounting height − ear height) × M, where M is 1.73 for centre-to-centre coverage on a 120° loudspeaker. But the bigger lever is the room. From RT 1.2 s at 10 dB signal-to-noise, adding 10 dB of level buys +0.04 STI while halving the reverberation time buys +0.11.

THE SENTENCE THE WHOLE FIELD GETS BACKWARDS

The metric is a method

Clause 1 of IEC 60268-16:2020+COR1:2025: it "does not provide STI criteria for certification of transmission channels (e.g. criteria for a voice-alarm system), but some typical application values are provided in Annex G".

WHAT THE NUMBER ACTUALLY IS

STI measures the rhythm that survives

Speech is loudness rising and falling a few times a second. The test signal carries that rhythm and the meter measures how much of it arrives: reverberation fills the dips, noise lifts the floor.

FULL STI AGAINST STIPA

STIPA reads 14 of the 98 cells

Full STI samples 14 modulation rates in each of 7 octave bands: 98 cells. STIPA samples two per band, 14 cells.

TWO STANDARDS, ONE MASHED-UP TABLE

The word scale is ISO 9921, the letters are IEC

IEC 60268-16's own scale is twelve letter bands, U to A+: ten of them 0.04 wide between 0.36 and 0.76, U and A+ open-ended, each quoted value a band centre. Bad to excellent is ISO 9921 's scale; IEC's informative annexes borrow the words and point there.

Reference

STI ranges × intelligibility class × use

Per IEC 60268-16:2020+COR1:2025, Annex F. These are the standard's own letter bands. The word labels Bad / Poor / Fair / Good / Excellent are ISO 9921, a separate ergonomics standard, and the letter "V" exists in neither scheme.

STI	Intelligibility · class	Spec required
Band U	below 0.36	Not suitable for PA systems (Annex G). The standard advises against quoting STI below 0.30 at all
Band J	0.36 to 0.40	Not suitable for PA systems (Annex G)
Band I	0.40 to 0.44	ISO 9921 would call this poor
Band H	0.44 to 0.48	Annex G's normal lower limit for voice alarm, nominal 0.46. Contains the NFPA 72 per-point floor of 0.45
Band G	0.48 to 0.52	Annex G's target value for voice alarm, nominal 0.50. Contains the NFPA 72 space average of 0.50
Band F	0.52 to 0.56	Annex G: good quality PA systems, as in shopping malls, public offices and cathedrals
Band E	0.56 to 0.60	Annex G example: modern churches, nominal 0.58
Band D	0.60 to 0.64	Annex G example: lecture rooms and classrooms, nominal 0.62
Band C	0.64 to 0.68	ISO 9921 would call this good
Band B	0.68 to 0.72	ISO 9921 would call this good
Band A	0.72 to 0.76	Annex G example: theatres, speech auditoria, parliaments and courts
Band A+	0.76 and above	Annex G example: recording studios, rarely achievable elsewhere

THE ROOM SETS A CEILING

In the reverberant field, the room sets the ceiling

Calculated, from the standard's own statistical model: a listener out in the reverberant field, no direct sound, one reverberation time (RT), no noise. STI 0.70 at RT 0.6 s, 0.59 at 1.0 s, 0.48 at 1.6 s and 0.30 at 4.0 s; 0.50 goes at RT 1.48 s, 0.45 at RT 1.88 s.

QUANTIFY THE THESIS

The room is nearly three times the lever

From RT 1.2 s at 10 dB signal-to-noise, STI 0.50: 10 dB more level buys +0.042, halving the reverberation time buys +0.115. 2.7 times the return, on the same room.

THE SAUDI DESIGN CASE

The emptiest prayer is the hardest room

A 30 × 30 × 8 m hall in marble and plaster rings for 12.5 s. Carpet takes it to 3.4 s.

THE THIRD LEVER

Aim the sound. Do not turn it up.

Direct sound falls 6 dB per doubling of distance; the reverberant field does not. Inside the critical distance Dc = 0.141 √(Q R) the listener beats that ceiling.

Deep Dive

Designing for STI, not for SPL

Five practical rules for PA in Saudi mosques, malls, and stations.

Reverberation kills STI more than noise does.

It can, and the model puts a number on it, so read the number with its assumptions. Calculated from the standard's statistical model for a listener out in the reverberant field, with no direct sound, one reverberation time and no noise: STI 0.70 at RT 0.6 s, 0.59 at 1.0 s, 0.48 at 1.6 s, 0.44 at 2.0 s and 0.30 at 4.0 s, so 0.50 goes at about RT 1.48 s and 0.45 at about 1.88 s. Those are that listener's limits, not a physical law: level alone cannot move them, as the Acoustical Society of America puts it, reinforcement makes speech louder and if it was unclear before it stays unclear, only louder. Directivity, placement, coverage and delay alignment can move them, by improving the direct-to-reverberant ratio.

Distributed beats centralised for STI.

Correct, and the reason is geometry, not power. Direct sound falls 6 dB per doubling of distance and the reverberant field does not, so what matters is how far each listener sits from a loudspeaker against the critical distance Dc = 0.141 √(Q R). Calculated from the standard's own Annex L, in a 30 × 30 × 8 m prayer hall with 300 worshippers (RT 2.3 s): a 120° cone (Q 3.7) holds STI 0.50 to about 10.6 m and a 120° × 30° column (Q 13.9) to about 20.6 m. The same hall rings for 12.5 s bare, 3.4 s carpeted and 1.4 s with 900 worshippers, which is a full floor at the published prayer density of about 0.96 m² per person. Treat that ladder as a worked illustration: it is Sabine with one broadband absorption coefficient per surface, published 500 Hz carpet values would run it about 1.7 times longer, and from about 380 worshippers the hall's mean absorption clears 0.2, where a diffuse field stops existing and both Sabine and Dc become extrapolations. A field study of 21 Saudi mosques found acoustic quality off optimum unoccupied and much better occupied, so the sparsely attended prayer is the design case.

Background noise is the second STI killer.

A working margin, yes, but note the ceiling on the other side, and note who says where it is. IEC 60268-16's model applies level-dependent auditory masking, in which each octave band masks the band above it, and that masking grows continuously with level: there is no knee at any one value in the standard. The nearest peer-reviewed figure is a measured one: word recognition in noise falls once speech passes about 69 dB SPL at a constant signal-to-noise ratio. The commonly quoted 80 dBA ceiling comes from meter documentation, not from the standard. NFPA 72 separately caps total sound pressure at 110 dBA at the minimum hearing distance.

Measure it quiet and add the crowd mathematically.

It does, and the standard's answer is the opposite of testing occupied: it provides for adding an occupancy noise spectrum to the measured modulation transfer matrix in post-processing, and NFPA 72 Annex D permits it. Be precise about what that adds. It adds the crowd's noise, not the crowd's absorption, which lowers the reverberation time and is a separate correction in the same clause list. What you must not do is measure in silence and report the number: the standard carries a clause titled overestimation of STI under low background noise conditions, and says a realistic background level is essential. Two site checks: set the test signal from the A-weighted equivalent level of at least 40 s of real speech at the position and add 3 dB, and run the analyser with the signal off, because a reading above about 0.2 means the background is fluctuating too much to trust.

In the Saudi documents we reviewed, no acceptance number.

We did not identify a universal numerical STI acceptance threshold in the documents reviewed. Confirm the applicable code edition, the project requirements and the requirements of the approving authority. Not found is not the same as does not exist, so here is exactly what was reviewed: SBC 801 of 2007 and of 2018 in full, the 2024 abridged Arabic SBC 201, the tables of contents of the 2024 volumes, 32 Civil Defence Council safety regulations, both Civil Defence submission guides, and SASO's fire-equipment technical regulation, which excludes fire alarm systems from its scope. We could not open the official English SBC volumes, where the referenced-standards chapters sit, so we make no claim about them. What the reviewed chain does say: SBC 201 hands installation and acceptance testing to NFPA 72, whose mandatory text makes quantitative measurement permitted but not required, and SBC 201 asks that the clarity of alarm signals be verified. An approving authority can add a requirement the code does not state. Our recommendation, as authors: where the project needs a number, write the criterion and its source document into the specification. GSO adopted ISO 7240-19 as a Gulf Standard in 2013 and IEC 60268-16:2020 in September 2026; a Gulf Standard governs a job once a technical regulation, an authority or the specification adopts it.

THE THRESHOLD IS TWO NUMBERS

≥0.45 at 90% of locations, ADS average ≥0.50

A room where every location reads 0.50 passes both parts. One averaging 0.54 with 15% of its locations at 0.40 passes the average and fails the 90% test.

ACCEPTANCE CRITERIA VARY BY APPLICATION AND JURISDICTION

There is no natural constant at 0.50

A 2008 US defence mass-notification document, inactive since September 2026, asks indoors for 0.7 CIS for the Navy and Marine Corps and 0.8 CIS for the Army and Air Force, each with its own rounding note.

THE SAUDI ANSWER

No acceptance number, in the documents we reviewed

SBC 201 and SBC 801 say where a voice alarm system is required and hand off to NFPA 72 for how to build and test it, at clauses 907-6 and 907-7 of the 2024 Arabic issue. NFPA 72's mandatory text then says quantitative measurements shall not be required, and shall be permitted.

THE GULF-NATIVE ROUTE NOBODY INVOKES

GSO already adopted the design standard

And the rule that is easiest to miss: no more than one third of points on a loudspeaker axis. Directly under a loudspeaker the direct sound flatters the reading, so a set of positions chosen for convenience records the best case rather than the space.

HOW A GOOD REPORT IS CHECKED

Three things that invalidate a STIPA result

One. Set the test signal from real speech: take the A-weighted equivalent level over at least 40 s of talker, or the whole recorded announcement, then add 3 dB, because the pauses pull that average below the speech itself.

THINGS THAT ARE STALE IN 2026

Dead documents still turn up in public address specifications

The %ALcons-to-STI equation built into major prediction software has, in the STI book editor's own words, no documented source in open literature.

THE GULF DETAILS

Two languages, one crowd, thirty degrees of drift

The standard puts numbers on the bilingual problem. For an intermediate second-language listener to receive the equivalent of 0.45, the system must measure 0.60; for an older listener with a 15 dB pure-tone average hearing loss, 0.57.

From the Academy to the Field

NLC, intelligibility specified with its threshold named.

We name the metric and its edition, name the threshold and the document it comes from, define the unit of assessment as the acoustically distinguishable space, set the point count per ISO 7240-19 with no more than one third of points on a loudspeaker axis, and state the reverberation time the room will be handed over at, because above about 1.5 s the target is unreachable at any level.

See the solution

Sources

Sources & further reading

The standards and codes behind every number on this page.

IEC 60268-16:2020 (edition 5.0) and IEC 60268-16:2011 (edition 4.0), the STI standard. Clause 1 of edition 5 provides no STI criteria for certification, pointing instead to the informative Annex G; edition 4 carries the Annex F bands, the Annex B STIPA table, the Annex J test-signal procedure, the Annex K report sheet and the normative Annex L prediction model. Clause numbers moved between editions, so each citation here names the edition it was read in. iec.ch
ISO 9921:2003, the ergonomics standard that the bad to excellent word scale comes from. IEC's own informative annexes borrow those words and point here for their definitions. iso.org
NFPA 72, Annex D and Chapter 18, the two-part criterion, 90% of locations at STI 0.45 and an average of 0.50, sits in Annex D. NFPA's Manual of Style marks every informational annex as not part of the requirements. The mandatory text (18.4.11 in the 2019 and 2022 editions, renumbered 18.4.12 in 2025) says quantitative measurements shall not be required and shall be permitted; the ADS definition at 3.3.6 is mandatory. nfpa.org
UFC 4-021-01, 9 April 2008 with Change 1, January 2010, the US defence mass-notification criteria: 0.7 CIS for the Navy and Marine Corps, 0.8 CIS for the Army and Air Force. Historical: WBDG lists it as inactive since 3 September 2026, and its individual-building rules were superseded by UFC 3-600-01. It states its own pairings (STI 0.5 equivalent to 0.7 CIS, STI 0.7 equivalent to 0.8 CIS) and allows 0.75 to round to 0.8; these are stated equivalences, not exact conversions (exact: 0.8 CIS is STI 0.63, and STI 0.70 is 0.845 CIS). wbdg.org
Garlinska, Popielarczyk, Sowa and Sliwinski, Safety and Fire Technology 57(1), 2021, the peer-reviewed comparison of voice alarm requirements across six European countries. Its only numeric STI values are Sweden's: 0.55 for evacuation messages and, in a different clause of the same building regulation, 0.60 facility-wide with 0.70 in more than half the rooms. doi.org
ISO 7240-19:2007, confirmed 2025, the design, installation, commissioning and service standard for emergency sound systems, adopted by GSO as a Gulf Standard in 2013. Its Annex A is the normative measurement method, and clause 5.7 covers speech intelligibility. The point counts and the one-third-on-axis rule quoted here are as reproduced in instrument-maker documentation; we have not read Annex A itself. iso.org
Saudi Building Code, SBC 201 and SBC 801, the Building Code is mandatory on the designer, the contractor and the supervisor under Royal Decree M/43 of 1438H, and its referenced standards are part of its requirements. SBC 201 hands installation and acceptance testing of the voice alarm system to NFPA 72 and asks that the clarity of alarm signals be verified. Neither volume we could open sets an STI number. sbc.gov.sa
A. A. Abdou, Journal of the Acoustical Society of America 113(3), 2003, field measurements in 21 representative Saudi mosques, finding acoustic quality off optimum unoccupied and much better occupied. The prayer density of about 0.96 m2 per person used in our worked hall comes from the same work, as applied by Su and Yilmazer in 2008. doi.org
S. van Wijngaarden (ed.), Past, present and future of the Speech Transmission Index, the modulation transfer function the STI is built from, the plus or minus 15 dB window, and the editor's note that the %ALcons to STI conversion built into prediction software has no documented source in open literature. embedded.engineering
Embedded Acoustics BA-WP-2020-199, with the TNO validation of STIPA meters, where Full STI is required instead of STIPA: distinct echoes, loudspeakers spaced far apart without delay lines, and reverberation above about 3 s with band limiting. The TNO comparison against the full 98-cell reference reported most data within 0.02 and a worst case of 0.03, stated as CIS in one sentence of the source and as STI in another. embedded.engineering
JBL Professional, Sound System Design Reference Manual, critical distance and the room constant, Molloy's equation for Q from coverage angles, and the statistical model's own limit: above a mean absorption of about 0.2 a diffuse reverberant field will not exist. JBL manual (PDF)
Acoustical Society of America, Classroom Acoustics for Architects, about 0.55 m2 of absorption per adult occupant, a classroom value applied here as an approximation, and the line that amplification leaves an unclear room unclear, only louder. acousticalsociety.org
Studebaker, Sherbecoe, McDaniel and Gwaltney, JASA 105(4), 1999, measured word recognition in noise falls once speech passes about 69 dB SPL at a constant signal-to-noise ratio, which is the nearest peer-reviewed figure to the 80 dBA ceiling quoted in instrument documentation. doi.org

Public Address, intelligibility, not loudness