5D Library › Architectural Acoustics, Reverberation Time and the Environmental Psychology of Built Space · 中文

Chapter 2 Standing Waves, Room Modes and Low-Frequency Noise

Where reverberation theory fails

The formulas of Sabine and Eyring in Chapter 1 both assume a “diffuse” sound field, in which many waves of similar frequency travelling in all directions coexist and their energy mixes evenly. At middle and high frequencies, and in larger rooms, this assumption roughly holds. At low frequencies in a small room, however, the room can sustain only a few stable patterns of vibration, its normal modes (room modes). Each mode is a standing wave formed by waves reflected back and forth between walls and adding together, with places of very high pressure (antinodes) and places of almost none (nodes). The low-frequency response of the room then depends on the particular frequency and position, and no longer on a single reverberation time.

Derivation: normal frequencies of a rectangular room

Calculation

For a rectangular room with rigid walls (length LL, width WW, height HH), the sound pressure satisfies the wave equation with zero normal particle velocity at the boundaries. Separating variables, the solutions that satisfy the boundary conditions are

pnxnynz(x,y,z)=cos⁡nxπxLcos⁡nyπyWcos⁡nzπzH,p_{n_x n_y n_z}(x, y, z) = \cos\frac{n_x\pi x}{L}\,\cos\frac{n_y\pi y}{W}\,\cos\frac{n_z\pi z}{H},

with normal frequencies

f=c2(nxL)2+(nyW)2+(nzH)2,nx,ny,nz=0,1,2,…f = \frac{c}{2}\sqrt{\left(\frac{n_x}{L}\right)^2 + \left(\frac{n_y}{W}\right)^2 + \left(\frac{n_z}{H}\right)^2},\qquad n_x, n_y, n_z = 0, 1, 2, \dots

Modes with one non-zero index are axial (between one pair of parallel walls), those with two are tangential and those with three are oblique. The number of modes below frequency ff is approximately

N(f)≈4πVf33c3+πSf24c2+Ef8c,N(f) \approx \frac{4\pi V f^3}{3c^3} + \frac{\pi S f^2}{4c^2} + \frac{E f}{8c},

where SS is the total surface area and E=4(L+W+H)E = 4(L + W + H) the total edge length (Kuttruff 2009). The count grows as f3f^3: sparse at low frequencies, dense at high ones.

When two or more modes fall at exactly the same frequency they are called degenerate. If the length, width and height are in simple integer ratios, many different combinations (nx,ny,nz)(n_x, n_y, n_z) give the same frequency, so modes “pile up” at a few frequencies and leave large gaps between them. Bolt discussed room proportions in 1946 in terms of the statistics of mode spacing (Bolt 1946); in 1981 Bonello proposed a widely used criterion: counting modes in third-octave bands, the number per band should never decrease with frequency (Bonello 1981).

The program below compares three rooms of equal volume (about 46.7 m³, a large study or a small meditation hall): a 3.6 m cube, a 1:2:2 room and the often recommended 1:1.4:1.9.

Calculation

# Rectangular-room modes f = (c/2) sqrt((nx/L)^2 + (ny/W)^2 + (nz/H)^2) for three rooms
# of equal volume (about 46.7 m3); degeneracy, gaps, Bonello's third-octave counts,
# Schroeder frequency.
import math, itertools

c = 343.0
ROOMS = {"cube 3.60 m": (3.60, 3.60, 3.60), "1:2:2": (4.54, 4.54, 2.27),
         "1:1.4:1.9": (4.94, 3.64, 2.60)}

def modes(L, W, H, fmax):
    out = []
    for n in itertools.product(range(12), repeat=3):
        f = c / 2 * math.sqrt((n[0] / L) ** 2 + (n[1] / W) ** 2 + (n[2] / H) ** 2)
        if 0 < f <= fmax:
            out.append((f, ("axial", "tangential", "oblique")[sum(k > 0 for k in n) - 1]))
    return sorted(out)

def weyl(L, W, H, f):              # expected number of modes below f (Weyl / Maa formula)
    V, S, E = L * W * H, 2 * (L * W + L * H + W * H), 4 * (L + W + H)
    return 4 * math.pi * V * f ** 3 / (3 * c ** 3) + math.pi * S * f ** 2 / (4 * c ** 2) \
        + E * f / (8 * c)

THIRDS = [20 * 2 ** (k / 3) for k in range(11)]     # third-octave edges 20 ... 200 Hz
for name, (L, W, H) in ROOMS.items():
    m = modes(L, W, H, 200.0)
    fs = [f for f, _ in m]
    distinct = sorted(set(round(f, 1) for f in fs))
    low = [f for f in distinct if f <= 120]
    gaps = [b - a for a, b in zip(low, low[1:])]
    kinds = {k: sum(x == k for _, x in m) for k in ("axial", "tangential", "oblique")}
    counts = [sum(1 for f in fs if a <= f < b) for a, b in zip(THIRDS, THIRDS[1:])]
    print(f"{name:10s} {L:.2f} x {W:.2f} x {H:.2f} m, V = {L * W * H:.1f} m3")
    print(f"   modes <= 200 Hz: {len(m)} ({kinds['axial']} axial, {kinds['tangential']}"
          f" tangential, {kinds['oblique']} oblique); Weyl {weyl(L, W, H, 200):.0f}")
    print(f"   distinct frequencies: {len(distinct)}; lowest {distinct[0]:.1f} Hz;"
          f" largest gap below 120 Hz {max(gaps):.1f} Hz")
    print("   modes per third-octave 20-200 Hz: " + " ".join(f"{n:2d}" for n in counts)
          + ("  (never decreasing)" if all(b >= a for a, b in zip(counts, counts[1:]))
             else "  (decreases)"))
for T in (0.4, 0.8):
    print(f"Schroeder frequency 2000 sqrt(T/V) for V = 46.7 m3, T = {T} s:"
          f" {2000 * math.sqrt(T / 46.7):.0f} Hz")
# cube 3.60 m 3.60 x 3.60 x 3.60 m, V = 46.7 m3
#    modes <= 200 Hz: 62 (12 axial, 30 tangential, 20 oblique); Weyl 63
#    distinct frequencies: 15; lowest 47.6 Hz; largest gap below 120 Hz 19.8 Hz
#    modes per third-octave 20-200 Hz:  0  0  0  3  0  3  4  9 18 25  (decreases)
# 1:2:2      4.54 x 4.54 x 2.27 m, V = 46.8 m3
#    modes <= 200 Hz: 62 (12 axial, 31 tangential, 19 oblique); Weyl 64
#    distinct frequencies: 20; lowest 37.8 Hz; largest gap below 120 Hz 22.2 Hz
#    modes per third-octave 20-200 Hz:  0  0  2  0  1  3  5  9 16 26  (decreases)
# 1:1.4:1.9  4.94 x 3.64 x 2.60 m, V = 46.8 m3
#    modes <= 200 Hz: 63 (12 axial, 31 tangential, 20 oblique); Weyl 64
#    distinct frequencies: 62; lowest 34.7 Hz; largest gap below 120 Hz 12.4 Hz
#    modes per third-octave 20-200 Hz:  0  0  1  1  1  3  6  7 18 26  (never decreasing)
# Schroeder frequency 2000 sqrt(T/V) for V = 46.7 m3, T = 0.4 s: 185 Hz
# Schroeder frequency 2000 sqrt(T/V) for V = 46.7 m3, T = 0.8 s: 262 Hz

The three rooms have almost the same number of modes below 200 Hz (62–63, in line with the approximation’s 63–64); the difference lies entirely in their distribution. The cube’s 62 modes fall on only 15 distinct frequencies; at 47.6 Hz, for example, three axial modes coincide. The 1:2:2 room has only 20 distinct frequencies; the 1:1.4:1.9 room’s 63 modes fall on 62. The largest gap between neighbouring modes below 120 Hz is 19.8 Hz for the cube, 22.2 Hz for 1:2:2 and only 12.4 Hz for 1:1.4:1.9, and only the last satisfies Bonello’s “never decreasing” criterion. In this sense the author’s statement that “a cubic room produces severe degeneracy and piling-up of low-frequency modes” is correct.

The Schroeder frequency

When are modes dense enough for the field to be treated as diffuse? Schroeder’s answer was: when, on average, about three other modes overlap within the resonance bandwidth of each mode. The half-power bandwidth of a mode is about 2.2/T2.2/T Hz; using the modal density dN/df≈4πVf2/c3dN/df \approx 4\pi V f^2/c^3 and setting their product equal to 3 gives

fs≈2000T/V(Hz; T in s, V in m3).f_s \approx 2000\sqrt{T/V}\quad(\text{Hz; } T\text{ in s, } V\text{ in m}^3).

The last two lines of the program give, for a 46.7 m³ room, fs≈185f_s \approx 185 Hz at a reverberation time of 0.4 s and about 262 Hz at 0.8 s. Below this frequency neither the reverberation formulas of Chapter 1 nor “clapping and listening to the tail” can be trusted; low-frequency boom is a problem of individual modes and has to be analysed frequency by frequency and position by position.

What is heard at a seat

Degeneracy stands out on the frequency axis, but a person sitting at one place hears the sum of all modes at that point. The next program uses a modal sum to compute the low-frequency response with the source in a corner and the listener at nine different seats, measures the irregularity by the standard deviation of the 50–150 Hz response, and compares two amounts of damping (a reverberation time of 0.8 s, and 0.4 s after adding low-frequency absorption). The model uses the same damping for every mode, a simplification.

Calculation

# Low-frequency response at one seat: modal sum p(w) ~ sum psi(src) psi(rcv) / (L_n
# (wn^2 - w^2 + 2j d w)), d = 6.91 / T for every mode (a simplification). Source near a
# corner; nine seated listening positions. Spread (SD) of the response 50-150 Hz in dB:
# smaller = smoother bass.
import math, itertools

c = 343.0
ROOMS = {"cube 3.60 m": (3.60, 3.60, 3.60), "1:1.4:1.9": (4.94, 3.64, 2.60)}

def mode_list(L, W, H, fmax=400.0):          # norm: factor 0.5 per non-zero index
    out = []
    for nx, ny, nz in itertools.product(range(12), repeat=3):
        f = c / 2 * math.sqrt((nx / L) ** 2 + (ny / W) ** 2 + (nz / H) ** 2)
        if 0 < f <= fmax:
            out.append((f, nx, ny, nz, 0.5 ** ((nx > 0) + (ny > 0) + (nz > 0))))
    return out

def response(dims, T, src, rcv, freqs):
    L, W, H = dims
    d = 6.91 / T
    ms = mode_list(L, W, H)
    shape = lambda n, x, D: math.cos(n * math.pi * x / D)
    psi = lambda nx, ny, nz, r: (shape(nx, r[0], L) * shape(ny, r[1], W)
                                 * shape(nz, r[2], H))
    weights = [(2 * math.pi * f, psi(nx, ny, nz, src) * psi(nx, ny, nz, rcv) / norm)
               for f, nx, ny, nz, norm in ms]
    out = []
    for f in freqs:
        w = 2 * math.pi * f
        p = sum(a / (wn * wn - w * w + 2j * d * w) for wn, a in weights)
        out.append(20 * math.log10(abs(p) * w + 1e-12))    # x w: ~ constant-volume source
    return out

def spread(db):
    m = sum(db) / len(db)
    return math.sqrt(sum((x - m) ** 2 for x in db) / len(db))

FREQS = [50 + 0.5 * k for k in range(201)]                   # 50-150 Hz
SEATS = [(fx, fy) for fx in (0.3, 0.45, 0.6) for fy in (0.3, 0.45, 0.6)]
for name, dims in ROOMS.items():
    L, W, H = dims
    src = (0.3, 0.3, 0.3)
    for T in (0.8, 0.4):
        sds = [spread(response(dims, T, src, (fx * L, fy * W, 1.2), FREQS))
               for fx, fy in SEATS]
        print(f"{name:11s} T = {T} s: SD over 9 seats mean {sum(sds) / 9:4.1f} dB"
              f" (best {min(sds):4.1f}, worst {max(sds):4.1f})")
# cube 3.60 m T = 0.8 s: SD over 9 seats mean  7.3 dB (best  5.8, worst 10.3)
# cube 3.60 m T = 0.4 s: SD over 9 seats mean  6.0 dB (best  4.6, worst  9.9)
# 1:1.4:1.9   T = 0.8 s: SD over 9 seats mean  7.6 dB (best  5.7, worst  9.3)
# 1:1.4:1.9   T = 0.4 s: SD over 9 seats mean  6.4 dB (best  4.9, worst  8.9)
Left: normal frequencies below 150 Hz for a cube and a 1:1.4:1.9 room of equal volume; stem height is the number of coincident modes at that frequency. Right: the 30–150 Hz response at one seat (modal-sum model, T = 0.8 s), each relative to its own mean.

The result surprises some people: in this simplified model the mean irregularity over nine seats is 7.3 dB for the cube and 7.6 dB for 1:1.4:1.9, almost the same; the difference between the best and worst seat within one room (5.8 versus 10.3 dB for the cube) is far larger than the difference between the two proportions; and lowering the reverberation time from 0.8 to 0.4 s reduces the irregularity in both rooms by about 1.2–1.3 dB. Figure 2 (right) shows that the cube’s problem is not that it is “worse everywhere”, but that its peaks are concentrated at a few frequencies with deeper and wider gaps between modes.

Note

Qualification of the author’s framework. Room proportions do determine how modes are distributed in frequency, but at a single seat the positions of listener and source and the low-frequency damping matter at least as much. Cox, D’Antonio and Avis argued in 2004 for optimising room dimensions together with loudspeaker and listener positions, allowing for construction tolerances (Cox et al. 2004). For a study or meditation hall that has already been built, moving the seat and adding low-frequency absorption in the corners is usually far more practical than changing the room’s proportions.

Low-frequency noise: a hum hidden by A-weighting

Air-handling units, fans, pumps, transformers and traffic all produce a great deal of low-frequency noise. The common A-weighted level, dB(A), mimics the ear’s sensitivity to sounds of moderate loudness and discounts low frequencies heavily: by about 39.5 dB at 31.5 Hz and about 26 dB at 63 Hz. C-weighting hardly discounts low frequencies at all.

Calculation

# Why dB(A) hides low-frequency noise: A- and C-weighting (IEC 61672 analytic forms)
# applied to two illustrative ventilation spectra adjusted to the same 40 dB(A), as in
# laboratory comparisons of a low-frequency-dominated and a flatter ventilation noise.
# Octave levels are invented.
import math

def r_a(f):
    f2 = f * f
    return 12194 ** 2 * f2 * f2 / ((f2 + 20.6 ** 2) * math.sqrt((f2 + 107.7 ** 2) *
                                    (f2 + 737.9 ** 2)) * (f2 + 12194 ** 2))
def r_c(f):
    f2 = f * f
    return 12194 ** 2 * f2 / ((f2 + 20.6 ** 2) * (f2 + 12194 ** 2))
A = lambda f: round(20 * math.log10(r_a(f)) + 2.00, 2) + 0.0
C = lambda f: round(20 * math.log10(r_c(f)) + 0.06, 2) + 0.0
BANDS = (31.5, 63, 125, 250, 500, 1000, 2000, 4000)
print("weighting (dB):  " + "  ".join(f"{b:>6}" for b in BANDS))
print("   A             " + "  ".join(f"{A(b):6.1f}" for b in BANDS))
print("   C             " + "  ".join(f"{C(b):6.1f}" for b in BANDS))
SPECTRA = {  # unweighted octave-band levels, dB (shape only; rescaled below)
    "flatter ventilation":   (48, 48, 46, 44, 40, 36, 32, 28),
    "low-frequency rumble":  (72, 70, 62, 50, 40, 34, 28, 22)}
total = lambda levels, w: 10 * math.log10(sum(10 ** ((L + w(b)) / 10) for L, b in
                                              zip(levels, BANDS)))
flat = lambda b: 0.0
for name, lv in SPECTRA.items():
    shift = 40.0 - total(lv, A)                             # rescale to exactly 40 dB(A)
    lv = [L + shift for L in lv]
    la, lc, lz = total(lv, A), total(lv, C), total(lv, flat)
    print(f"{name:21s}: {la:.1f} dB(A), {lc:.1f} dB(C), {lz:.1f} dB unweighted;"
          f" C - A = {lc - la:4.1f} dB")
    if lc - la >= 20:
        print(f"{'':23s}C - A >= 20 dB: low-frequency noise by the DIN 45680 rule")
    share = 10 ** ((lv[0] + A(31.5)) / 10) + 10 ** ((lv[1] + A(63)) / 10)
    print(f"{'':23s}share of the A-weighted energy below 90 Hz:"
          f" {share / 10 ** (la / 10):.0%}; unweighted: "
          f"{(10 ** (lv[0] / 10) + 10 ** (lv[1] / 10)) / 10 ** (lz / 10):.0%}")
# weighting (dB):    31.5      63     125     250     500    1000    2000    4000
#    A              -39.5   -26.2   -16.2    -8.7    -3.2     0.0     1.2     1.0
#    C               -3.0    -0.8    -0.2     0.0     0.0     0.0    -0.2    -0.8
# flatter ventilation  : 40.0 dB(A), 50.1 dB(C), 51.1 dB unweighted; C - A = 10.1 dB
#                        share of the A-weighted energy below 90 Hz: 1%; unweighted: 61%
# low-frequency rumble : 40.0 dB(A), 63.2 dB(C), 65.1 dB unweighted; C - A = 23.2 dB
#                        C - A >= 20 dB: low-frequency noise by the DIN 45680 rule
#                        share of the A-weighted energy below 90 Hz: 30%; unweighted: 94%

Two ventilation noises are set to the same 40 dB(A). The flatter one has a C-weighted level of 50.1 dB, C − A about 10 dB; the one dominated by low frequencies has 63.2 dB(C), C − A of 23.2 dB, and an unweighted total 14 dB higher. Of the A-weighted energy, the part below about 90 Hz is 1 % for the first and 30 % for the second; of the unweighted energy it is 61 % and 94 %. The German standard DIN 45680 uses a rule of thumb: if the C-weighted level exceeds the A-weighted level by 20 dB or more, the noise should be assessed separately as low-frequency noise (Deutsches Institut für Normung 1997). An air conditioner “of only 40 decibels” can thus be producing a strong low-frequency hum that an ordinary sound level reading does not reveal.

Evidence

Evidence on low-frequency noise and physiological stress. Leventhall’s 2004 review concluded that annoyance from low-frequency noise (roughly 10–200 Hz) is real, that A-weighting underestimates it, and that sensitivity to it varies widely between people (Leventhall 2004). The 2002 study by Persson Waye and colleagues cited in the author’s framework is an experiment, not a review: 32 participants, classified by questionnaire as of high or low sensitivity to noise (to noise in general and to low-frequency noise), performed a battery of tasks for about 2 hours on each of two afternoons under two ventilation noises, both at 40 dB(A): one a flat ventilation noise with low-frequency content at 31.5–125 Hz added, the other the flatter noise alone; salivary cortisol was measured. Overall, cortisol levels were not significantly affected by noise type; only in participants highly sensitive to noise in general, exposed to the low-frequency noise, was the normal afternoon decline in cortisol significantly attenuated (grouping by sensitivity to low-frequency noise gave only a trend). On performance, only two tasks were worse under the low-frequency noise, in the second half of the session: fewer marks per line in proof-reading and longer response times in verbal grammatical reasoning; simple reaction time and short-term memory did not differ. Participants also rated the low-frequency noise as more annoying and more disruptive to their work (Persson Waye et al. 2002).

Note

Correction to the author’s framework. “Low-frequency noise from air conditioning and ventilation (20–100 Hz) impairs sustained attention and cortisol recovery (review, Waye et al. 2002)” needs rewriting: the paper is an experimental study of 32 people; the cortisol effect appeared only in the subgroup of highly noise-sensitive participants, as an attenuation of the normal decline; and performance was worse on only two tasks (proof-reading and grammatical reasoning) in the second half of the session, with no difference on the others. This is preliminary evidence worth taking seriously, but it cannot be summarised as “low-frequency noise impairs everyone’s sustained attention and stress recovery”. Moreover, the normal modes of a room (a matter of acoustic quality) and sources of low-frequency noise (mechanical plant) are different things: the former amplify or suppress certain frequencies, the latter supply the energy. No study was found showing directly that modal degeneracy causes physiological stress.

Hypothesis · not established

In small retreat or meditation rooms, low-frequency plant noise that coincides with a room mode is more often reported as “oppressive” than broadband noise of the same A-weighted level (not established). A testable consequence: playing two noises at the same 35 dB(A) in the same room, one with its energy concentrated at the frequency of a strong mode of that room and one between modes, blinded participants should rate the first as more oppressive, and the difference should remain when the C − A difference measured by a sound level meter is the same. If ratings follow C − A alone, regardless of whether the noise falls on a mode, the hypothesis is refuted.

Chapter summary

The normal frequencies of a rectangular room are given by f=c2(nx/L)2+(ny/W)2+(nz/H)2f = \frac{c}{2}\sqrt{(n_x/L)^2 + (n_y/W)^2 + (n_z/H)^2}, and the number of modes grows as f3f^3. At equal volume, a cube’s 62 modes below 200 Hz fall on only 15 frequencies, whereas a 1:1.4:1.9 room has almost no coincidences; the Schroeder frequency 2000T/V2000\sqrt{T/V} marks the boundary between the modal and diffuse regions (about 185–262 Hz for the example room). At a single seat, though, listener position and low-frequency damping matter at least as much as proportions. A-weighting hides low-frequency noise: two ventilation noises at the same 40 dB(A) can differ by 13 dB in C − A. The 2002 experiment by Persson Waye and colleagues suggests that low-frequency noise may affect the cortisol rhythm of highly sensitive people, but the evidence is limited. The next chapter discusses a property of sound that matters more than decibels: the voices in the background.

References

Bolt, Richard H. 1946. “Note on Normal Frequency Statistics for Rectangular Rooms.” The Journal of the Acoustical Society of America 18 (1): 130–33. https://doi.org/10.1121/1.1916349.
Bonello, Oscar J. 1981. “A New Criterion for the Distribution of Normal Room Modes.” Journal of the Audio Engineering Society 29 (9): 597–606.
Cox, Trevor J., Peter D’Antonio, and Mark R. Avis. 2004. “Room Sizing and Optimization at Low Frequencies.” Journal of the Audio Engineering Society 52 (6): 640–51.
Deutsches Institut für Normung. 1997. DIN 45680:1997 Messung Und Bewertung Tieffrequenter Geräuschimmissionen in Der Nachbarschaft. Beuth, Berlin.
Kuttruff, Heinrich. 2009. Room Acoustics. 5th ed. Spon Press.
Leventhall, H. Geoff. 2004. “Low Frequency Noise and Annoyance.” Noise and Health 6 (23): 59–72.
Persson Waye, Kerstin, Johanna Bengtsson, Ragnar Rylander, Frank Hucklebridge, Phil Evans, and Angela Clow. 2002. “Low Frequency Noise Enhances Cortisol Among Noise Sensitive Subjects During Work Performance.” Life Sciences 70 (7): 745–58. https://doi.org/10.1016/S0024-3205(01)01450-3.

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).