5D Library › Architectural Acoustics, Reverberation Time and the Environmental Psychology of Built Space · 中文
Chapter 3 Open-Plan Offices and the Irrelevant Speech Effect
A question that matters more than decibels
The commonest complaint in open-plan offices is not “it is too loud” but “someone nearby is talking”. An air conditioner and a colleague’s phone call at the same 45 dB(A) affect work very differently. Cognitive psychology has a name for this: the irrelevant speech effect, later more broadly called the irrelevant sound effect.
Evidence
Classic experiments. In 1976 Colle and Welsh asked participants to remember visually presented sequences of letters and recall them in order while foreign speech they did not understand was played; even though it was entirely incomprehensible, the background speech significantly increased errors in serial recall (Colle and Welsh 1976). In 1982 Salamé and Baddeley found that background speech which participants were told to ignore disrupted serial recall of visually presented digits, that “nonsense words” made of meaningless syllables disrupted it as much as real words, and that the semantic content of the background speech made no difference; they explained the result with the “phonological loop” of the working-memory model: speech sounds gain automatic access to the phonological store and interfere with the items being rehearsed (Salamé and Baddeley 1982).
Understanding, or change?
The author’s framework stated that what really breaks working memory is “intelligible speech by others”, because the phonological loop is forced to decode intelligible speech. The two classic experiments already suggest that understanding is not necessary, and later research took the question further.
Evidence
The changing-state hypothesis. Jones and Macken found in 1993 that sequences of tones of changing pitch produce a similar, though weaker, disruption; they argued that what matters is that the sound is segmented in time into successive, distinct “states”, and that this change itself disrupts memory for order, whether or not the sound is speech (Jones and Macken 1993). Ellermeier and Hellbrück found in 1998 that over a wide range the playback level of the background speech hardly affects the size of the disruption (Ellermeier and Hellbrück 1998). Ellermeier and Zimmer’s 2014 review summarised: semantics seem unimportant, and unfamiliar and native languages disrupt about equally; time-reversed speech still disrupts considerably; noticeable spectral change over time is a necessary condition for the effect, while level change alone is not; on average background speech increases serial-recall errors by about half, with very large individual differences (Ellermeier and Zimmer 2014). The 2001 review by Banbury and colleagues discusses these phenomena and their implications for the workplace (Banbury et al. 2001).
The program below brings together several kinds of sound listed in Ellermeier and Zimmer’s review, with their absolute increase in error rate over quiet (in percentage points), and marks whether each can be understood by the listener and whether it changes spectrally.
Calculation
# Part 1: which background sounds disrupt serial recall? Absolute increases in error
# rate (percentage points vs quiet) as tabulated in the review by Ellermeier & Zimmer
# (2014); compared with what an "understanding" account and a "changing-state" account
# would predict. Part 2: an illustrative STI -> serial-recall error curve (shape of
# Hongisto 2005, rescaled).
import math
DATA = [ # sound, understood?, spectral change over time?, ADER low, ADER high
("native speech", True, "yes", 8.8, 29.0),
("nonsense words", False, "yes", 8.8, 8.8),
("time-reversed speech", False, "yes", 6.0, 14.1),
("tone sequence, varying", False, "partial", 5.5, 5.5),
("level change only", False, "no", 2.2, 2.2),
("amplitude-modulated noise", False, "no", 1.2, 1.2),
("steady noise", False, "no", 0.0, 0.0)]
print(f"{'background':27s} understood changing-state ADER (pp vs quiet)")
for name, und, chg, lo, hi in DATA:
rng = f"{lo:4.1f}" if lo == hi else f"{lo:4.1f}-{hi:4.1f}"
rng = " ~0 (no reliable effect)" if hi == 0 else rng
print(f"{name:27s} {'yes' if und else 'no':9s} {chg:13s} {rng}")
not_und = [d for d in DATA if not d[1]]
print("not understood but changing-state: ADER "
+ ", ".join(f"{d[3]:.1f}" for d in not_und if d[2] == "yes")
+ " pp -> understanding is NOT needed")
print("no spectral change: ADER at most "
f"{max(d[4] for d in DATA if d[2] == 'no'):.1f} pp -> spectral change is the key")
E_QUIET, ADER_MAX = 30.0, 15.0 # illustrative: 30 % errors in quiet, max +15 pp
shape = lambda s: 1 / (1 + math.exp((0.4 - s) / 0.06))
STIS = (0.75, 0.65, 0.50, 0.40, 0.30, 0.25)
print("STI of background speech: " + " ".join(f"{s:5.2f}" for s in STIS))
print("predicted errors (%): "
+ " ".join(f"{E_QUIET + ADER_MAX * shape(s):5.1f}" for s in STIS))
d = ADER_MAX * (shape(0.65) - shape(0.25))
print(f"illustrative difference, STI 0.65 vs 0.25 at the same level: {d:.1f} pp")
print(f" (with ADER_MAX 8.8 instead of 15: {8.8 * (shape(0.65) - shape(0.25)):.1f} pp)")
# background understood changing-state ADER (pp vs quiet)
# native speech yes yes 8.8-29.0
# nonsense words no yes 8.8
# time-reversed speech no yes 6.0-14.1
# tone sequence, varying no partial 5.5
# level change only no no 2.2
# amplitude-modulated noise no no 1.2
# steady noise no no ~0 (no reliable effect)
# not understood but changing-state: ADER 8.8, 6.0 pp -> understanding is NOT needed
# no spectral change: ADER at most 2.2 pp -> spectral change is the key
# STI of background speech: 0.75 0.65 0.50 0.40 0.30 0.25
# predicted errors (%): 45.0 44.8 42.6 37.5 32.4 31.1
# illustrative difference, STI 0.65 vs 0.25 at the same level: 13.6 pp
# (with ADER_MAX 8.8 instead of 15: 8.0 pp)Sounds that cannot be understood but do change spectrally (nonsense words 8.8 points, reversed speech 6.0–14.1 points) all disrupt clearly; sounds without spectral change (level change alone, amplitude-modulated noise, steady noise) disrupt by at most 2.2 points. “Understanding” does not explain this table; “spectral change” does.
Note
Correction to the author’s framework. For tasks such as serial recall that depend on memory for order, disruption does not require understanding: foreign speech, nonsense words and reversed speech all produce it, and level is largely irrelevant. “The phonological loop is forced to decode intelligible speech” should be rewritten as: sounds with marked spectral change over time (speech being the most typical) gain automatic access to auditory processing and disrupt the retention of order information. For tasks such as reading comprehension that depend on meaning, however, meaning does matter: Martin, Wogalter and Forlano found in 1988 that meaningful background speech impaired reading comprehension, whereas meaningless speech did little more than white noise, and instrumental music did no harm (Martin et al. 1988). Two tasks, two mechanisms; they should not be conflated.
From the laboratory to the office: STI and performance
Laboratory studies compare “speech” with “no speech”; in an office the neighbour’s voice is always there, only more or less clear. Hongisto connected this with the speech transmission index of Chapter 1.
Evidence
The Hongisto model. Pooling several laboratory studies, Hongisto proposed in 2005 that the decrement in work performance (relative to quiet) follows an S-shaped function of the STI of the background speech:
with almost no decrement below an STI of about 0.2, half the maximum at 0.4, and a ceiling of about 7 % above 0.6–0.7 (Hongisto 2005). The model works for partially masked or poorly transmitted speech, but it requires knowing in advance that the background is speech; for non-speech background sounds it does not give correct predictions (Ellermeier and Zimmer 2014).
Evidence
Field data and masking. ISO 3382-3 (2012) specifies how to measure open-plan offices; its distraction distance is the distance from a talker at which STI falls below 0.50, and its privacy distance the distance at which it falls below 0.20 (International Organization for Standardization 2012). Haapakangas and colleagues in 2017 pooled questionnaires from 883 employees in 21 open-plan offices with acoustic measurements made to ISO 3382-3, and found that a longer distraction distance predicted more reported disturbance by noise, whereas the spatial decay rate of speech, the speech level at 4 m and the background noise level did not, on their own, predict it (Haapakangas et al. 2017). In an experiment with 54 participants, Haapakangas and colleagues compared speech (STI 0.62), silence and five masking sounds (masked STI 0.38): compared with silence, short-term memory was worse with speech and with most masked conditions; spring-water sound was the best masker and improved performance relative to speech alone; all masking sounds raised satisfaction with the acoustic environment, while vocal music was as harmful as speech (Haapakangas et al. 2011).
Following ISO 3382-3, the program below builds four offices: the speech level is at 4 m and falls by dB per doubling of distance; the background (ventilation or masking) is dB(A); STI comes from the simplified formula of Chapter 1 ( = 0.6 s). In office 4 the masking is louder, and following the Lombard effect talkers are assumed to raise their voices by 1.5 dB as a result (0.5 dB per dB of background above 45 dB(A), an illustrative value); the Lombard effect itself is covered in Chapter 3 of Book 10 and Chapter 6 of Book 29 of this series.
Calculation
# Open-plan office in the style of ISO 3382-3: speech level falls by D2S dB per doubling
# of distance from L4m (A-weighted speech level at 4 m); background (ventilation or
# masking) B dB(A). STI from the simplified MTF of p02 with T = 0.6 s and speech-shaped
# noise (same SNR in every band). Hongisto (2005): performance decrement DP = 7 % / (1 +
# exp((0.4 - STI) / 0.06)).
import math
FM = [10 ** (k / 10) for k in range(-2, 12)]
ALPHA = (0.085, 0.127, 0.230, 0.233, 0.309, 0.224, 0.173)
BETA = (0.085, 0.078, 0.065, 0.011, 0.047, 0.095)
def sti(T, snr):
tis = []
for F in FM:
m = 1 / math.sqrt(1 + (2 * math.pi * F * T / 13.8) ** 2) / (1 + 10 ** (-snr / 10))
tis.append((max(-15.0, min(15.0, 10 * math.log10(m / (1 - m)))) + 15) / 30)
M = sum(tis) / len(tis)
return sum(ALPHA) * M - sum(BETA) * M # same MTI in all seven bands
dp = lambda s: 7.0 / (1 + math.exp((0.4 - s) / 0.06)) # % decrement vs silence
OFFICES = { # D2S dB/doubling, L4m dB(A), background dB(A), Lombard rise of talker (dB)
"1 hard, no screens": (5.0, 52.0, 35.0, 0.0),
"2 absorbent + screens": (8.0, 47.0, 35.0, 0.0),
"3 = 2 + masking 42 dB": (8.0, 47.0, 42.0, 0.0),
"4 = 2 + masking 48 dB": (8.0, 47.0, 48.0, 1.5)}
DIST = (1, 2, 4, 8, 16)
def at(office, r):
d2s, l4, bg, lomb = OFFICES[office]
s = sti(0.6, l4 + lomb + d2s * math.log2(4 / r) - bg)
return s, dp(s)
def radius(office, target):
r = 0.5
while at(office, r)[0] >= target and r < 40:
r *= 1.01
return f"{r:4.1f} m" if r < 40 else "> 40 m"
print("STI and predicted decrement DP at distance r from a talker")
print(f"{'office':23s}" + "".join(f"{r:>4d} m " for r in DIST))
for o in OFFICES:
print(f"{o:23s}" + "".join(f"{at(o, r)[0]:.2f} ({at(o, r)[1]:3.1f}%) " for r in DIST))
for o in OFFICES:
print(f"{o:23s} distraction rD (STI < 0.5) {radius(o, 0.5)},"
f" privacy rP (STI < 0.2) {radius(o, 0.2)}")
print("Hongisto model, STI -> DP: "
+ ", ".join(f"{s:.2f} -> {dp(s):.2f} %" for s in (0.7, 0.5, 0.3)))
print(" "
+ ", ".join(f"{s:.2f} -> {dp(s):.2f} %" for s in (0.25, 0.2)))
# STI and predicted decrement DP at distance r from a talker
# office 1 m 2 m 4 m 8 m 16 m
# 1 hard, no screens 0.70 (7.0%) 0.69 (6.9%) 0.68 (6.9%) 0.64 (6.9%) 0.57 (6.6%)
# 2 absorbent + screens 0.70 (7.0%) 0.69 (6.9%) 0.64 (6.9%) 0.51 (6.0%) 0.29 (1.0%)
# 3 = 2 + masking 42 dB 0.69 (6.9%) 0.65 (6.9%) 0.53 (6.3%) 0.32 (1.5%) 0.08 (0.0%)
# 4 = 2 + masking 48 dB 0.68 (6.9%) 0.59 (6.7%) 0.42 (4.1%) 0.18 (0.2%) 0.00 (0.0%)
# 1 hard, no screens distraction rD (STI < 0.5) 25.5 m, privacy rP (STI < 0.2) > 40 m
# 2 absorbent + screens distraction rD (STI < 0.5) 8.3 m, privacy rP (STI < 0.2) 20.9 m
# 3 = 2 + masking 42 dB distraction rD (STI < 0.5) 4.5 m, privacy rP (STI < 0.2) 11.4 m
# 4 = 2 + masking 48 dB distraction rD (STI < 0.5) 3.1 m, privacy rP (STI < 0.2) 7.7 m
# Hongisto model, STI -> DP: 0.70 -> 6.95 %, 0.50 -> 5.89 %, 0.30 -> 1.11 %
# 0.25 -> 0.53 %, 0.20 -> 0.24 %In office 1, hard, unscreened and with a quiet background, the STI is still 0.57 at 16 m; the distraction distance is about 25.5 m and the privacy distance more than 40 m: the whole floor is “listening in”. An absorbent ceiling with screens (office 2) brings the distraction distance down to 8.3 m; adding masking at 42 dB(A) (office 3) gives 4.5 m and a privacy distance of 11.4 m; raising the masking to 48 dB(A) (office 4) brings the distraction distance down to 3.1 m, even allowing for talkers raising their voices. But look at the first two columns of the table: at 1–2 m, the STI is between 0.59 and 0.70 in all four offices. Masking cannot protect the desk next to yours; speech at close range can be dealt with only by distance, screens, rules and quiet rooms.
In Hongisto’s model an STI of 0.70 corresponds to a decrement of about 6.95 %, 0.50 to about 5.9 %, 0.30 to about 1.1 % and 0.25 to about 0.5 %. The author’s claim that “errors fall significantly when STI drops from 0.7 to below 0.3” is right in direction, but the size has to be stated: this is a relative decrement against quiet, pooled over many tasks, with a ceiling of about 7 %. For the most sensitive tasks, such as serial recall, laboratory effects are larger: the second part of the earlier program uses the same S-shape with a maximum increase in serial-recall errors of 15 percentage points (between the 8.8 and 29 points of speech effects in the review, an illustrative value), giving a difference of about 13.6 points between STIs of 0.65 and 0.25; with a maximum of only 8.8 points, the difference is about 8.0 points. The trial in Chapter 8 is designed to measure this difference.
Note
An easily confused point. STI is a physical index of the transmission channel: in the same room, foreign speech and native speech have the same STI. It probably predicts the irrelevant speech effect not because it measures “how much is understood” but because it measures how much of the fluctuation of the speech envelope survives, which is the physical basis of “changing state”. Masking reduces both, so masking helps whichever explanation is right; but experiments that only lower STI cannot tell the two mechanisms apart.
Hypothesis · not established
The effects of masking on serial recall and on reading comprehension come from different mechanisms (not established). A testable consequence: at equal STI, replacing the background speech with a language participants do not understand should leave the disruption of serial recall essentially unchanged while clearly reducing the disruption of reading comprehension, and masking should reduce the disruption of serial recall equally in both language conditions. If the unknown language impairs reading comprehension just as much, or if native and foreign speech disrupt serial recall very differently, the hypothesis is weakened.
Chapter summary
The irrelevant speech effect is one of the most robust phenomena in cognitive psychology: background speech increases serial-recall errors by about half on average, almost regardless of level. But it does not require understanding: foreign speech, nonsense words and reversed speech all produce it, and the key is spectral change over time; for reading comprehension, meaning is what matters. At the scale of an office, Hongisto’s model writes the performance decrement as an S-shaped function of STI (with a ceiling of about 7 %), and the distraction distance of ISO 3382-3 predicts how disturbed employees are. In the illustrative calculation, absorption, screens and masking shrink the distraction distance from about 25 m to 3–5 m, but do almost nothing for neighbours within 1–2 m. The next chapter turns from the ear to the eye: what is seen through the window.
References
Fan, Y. (2026). Architectural Acoustics, Reverberation Time and the Environmental Psychology of Built Space: From Sabine's Equation and Fractal Visual Dimensions to Cognitive Performance, Read Beside Classical Chinese Spatial Design [建筑声学、混响时间与空间环境心理学的物理模型] (Version 1.0.0). 5D Library. https://doi.org/10.5281/zenodo.23782090 · Licensed under Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0).