Integrated and Statistical Levels #
Environmental noise metrics computed directly from the raw (calibrated) signal.
Leq and LAeq #
The equivalent continuous level integrates the squared pressure over the measurement time:
$$ L_{eq} = 10\log_{10}\left(\frac{1}{T}\int_0^T \frac{p^2(t)}{p_0^2}\ dt\right) \text{ dB}, \qquad p_0 = 20\ \mu\text{Pa} $$
and $L_{Aeq}$ is the same integral after A-weighting the signal. $L_N$ is the level exceeded $N\ %$ of the time — the $(100-N)$-th percentile of the time-weighted level distribution.
import numpy as np
from phonometry import leq, laeq
# recording: a calibrated microphone capture (Pa) — recorded through your measurement chain. Synthesized here so the guide runs standalone.
fs = 48000
recording = 0.2 * np.sin(2 * np.pi * 1000 * np.arange(fs) / fs)
sensitivity = 1.0 # calibration_factor (see Calibration)
# Equivalent continuous level of the whole recording
level = leq(recording, calibration_factor=sensitivity)
# A-weighted Leq (the standard environmental noise metric)
la = laeq(recording, fs, calibration_factor=sensitivity)
Both accept 1D signals (returning a scalar) or 2D [channels, samples] arrays
(returning one level per channel), and support dbfs=True for digital
full-scale analysis (calibration does not apply in dBFS mode).
Why the energy mean and not the arithmetic mean of dB values? Because sound doses add as energy: two periods at 60 dB and 80 dB do not average to 70 dB — the 80 dB half dominates and $L_{eq}$ = 77 dB. Averaging decibels directly underestimates every fluctuating noise. $L_{eq}$ is the level of the steady sound carrying the same energy as the real, fluctuating one, which is why regulations are written in terms of it.
leq() / laeq() parameters #
| Parameter | Type / shape | Units | Range / default | Notes |
|---|---|---|---|---|
x |
1D or 2D array | digital units (or Pa if calibrated) | non-empty | 2D is [channels, samples]; returns one level per channel |
fs |
int | Hz | > 0 (laeq only) |
leq needs no sample rate (pure RMS integral) |
calibration_factor |
float | Pa per digital unit | default 1.0 |
From calculate_sensitivity() |
dbfs |
bool | — | default False |
True: 0 dBFS = full-scale RMS sine; ignores calibration |
Percentile levels (LN) #
ln_levels computes statistical levels from the time-weighted envelope:
L10 is the level exceeded 10 % of the time (event peaks), L50 the median,
L90 the background level.
from phonometry import ln_levels
# A steady tone gives L10 = L50 = L90; percentiles only tell a story for a
# *fluctuating* level. Synthesize 3 s alternating between a quiet and a
# ~10 dB louder half-second so the statistics separate.
rng = np.random.default_rng(0)
segment = fs // 2 # 0.5 s per level
quiet = 0.02 * rng.standard_normal(segment) # background
loud = 0.06 * rng.standard_normal(segment) # ~10 dB louder events
varying = np.tile(np.concatenate([quiet, loud]), 3)
stats = ln_levels(varying, fs, n=(10, 50, 90), weighting="A")
print(f"LA10={stats[10]:.1f} LA50={stats[50]:.1f} LA90={stats[90]:.1f} dB")
# LA10=66.6 LA50=65.2 LA90=58.5 dB -> L10 (events) > L50 (median) > L90 (background)
L10 tracks the event peaks, L50 the median level and L90 the background.
Options: mode selects the envelope ballistics ('fast', 'slow',
'impulse'), weighting applies A/C weighting first, and
calibration_factor/dbfs behave as in leq. The integrator attack transient
(~5τ) is discarded before taking percentiles, so the leading settling ramp is
not counted in the low percentiles.
Formally, $L_N$ is the $(100-N)$-th percentile of the distribution of the time-weighted level: the recording is first turned into a level-vs-time envelope (Fast by default), and $L_{10}$ is the envelope value exceeded 10 % of the time. That makes the ballistics choice part of the metric: an $L_{10}$ from a Slow envelope is systematically lower than from a Fast one on impulsive noise, so regulations always name the time weighting.
ln_levels() parameters #
| Parameter | Type / shape | Units | Range / default | Notes |
|---|---|---|---|---|
x |
1D or 2D array | digital units | non-empty | 2D returns per-channel dicts |
fs |
int | Hz | > 0 | Needed by the envelope detector |
n |
tuple of ints | % | default (10, 50, 90) |
Any exceedance percentages, e.g. (1, 5, 95) |
mode |
str | — | 'fast' (default), 'slow', 'impulse' |
IEC 61672-1 ballistics of the envelope |
weighting |
str or None | — | 'A', 'C', 'G', 'Z', None (default) |
Frequency weighting before the envelope |
calibration_factor / dbfs |
float / bool | — | as leq |
Same semantics as in leq() |
Peak, event and occupational metrics #
from phonometry import lc_peak, sel, sound_exposure, lex_8h
# C-weighted peak (IEC 61672-1 §5.13) - occupational action limits use this
peak = lc_peak(recording, fs, calibration_factor=sensitivity)
# A single noise event and a work-shift sample (slices of a real recording)
event = recording
shift_sample = recording
# Sound exposure level: single-event level normalized to 1 s (LAE)
lae = sel(event, fs, weighting="A", calibration_factor=sensitivity)
# Daily noise dose (IEC 61252): exposure in Pa²·h and LEX,8h / LEP,d
E = sound_exposure(shift_sample, fs, duration_hours=8, calibration_factor=sensitivity)
lex = lex_8h(shift_sample, fs, duration_hours=8, calibration_factor=sensitivity)
lc_peak is verified against the one-cycle/half-cycle reference responses of
IEC 61672-1:2013 Table 5, sel against the Table 4 LAE toneburst column, and
the dose functions against the IEC 61252 anchors (3.2 Pa²h ↔ exactly 90 dB).
lc_peak polyphase-oversamples the C-weighted signal by oversample (default
8) before taking the maximum, recovering the true inter-sample peak: a raw
on-grid maximum under-reads sustained HF tones by up to ~1.15 dB (an 8 kHz tone
at 48 kHz is only 6 samples/cycle). Set oversample=1 to detect the peak on the
original sample grid. With duration_hours, the input is treated as a
representative sample of that exposure period; without it, the input is the
whole event.
SEL: comparing events of different duration #
A 4 s train pass-by and a 30 s one cannot be compared by their $L_{Aeq}$ alone — the longer event delivers more energy at the same level. The sound exposure level compresses the whole event energy into exactly one second:
$$ L_E = L_{eq,T} + 10\log_{10}\frac{T}{T_0}, \qquad T_0 = 1\ \text{s} $$
so events of any duration become directly comparable, and $N$ identical events sum as $+10\log_{10}N$. This is the building block of airport and railway noise models.
Noise dose: sound exposure and LEX,8h #
Occupational regulations limit the daily dose, not the level. IEC 61252 expresses it as sound exposure $E$ in pascal-squared-hours — the time integral of the squared A-weighted pressure — and the equivalent normalized 8 h level:
$$ E = \int_0^T p_A^2(t)\ dt \quad [\text{Pa}^2\text{h}], \qquad L_{EX,8h} = 10\log_{10}\frac{E}{8\ \text{h} \cdot p_0^2} $$
The anchor worth memorizing: 3.2 Pa²h ⇔ exactly 90 dB over 8 h (the CI suite enforces it). Half the dose is −3 dB; double duration at the same level is +3 dB.
Peak / event / dose parameters #
| Function | Key parameters | Returns | Standard anchor |
|---|---|---|---|
lc_peak(x, fs, calibration_factor=1.0, dbfs=False) |
dbfs=True references full-scale peak (1.0), not RMS |
LCpeak [dB] | IEC 61672-1 §5.13, Table 5 tone bursts |
sel(x, fs, weighting=None, ...) |
weighting='A' gives LAE |
SEL [dB] | IEC 61672-1 Table 4 (LAE column) |
sound_exposure(x, fs, duration_hours=None, ...) |
duration_hours treats x as a sample of that period |
E [Pa²h] | IEC 61252 |
lex_8h(x, fs, duration_hours=None, ...) |
same sampling semantics | LEX,8h [dB] | IEC 61252 (≡ LEP,d) |
Occupational noise exposure strategies and uncertainty (ISO 9612) #
lex_8h above turns one recording into a daily level. ISO 9612:2009 — the
engineering method (accuracy grade 2) — is the survey design around that
primitive: how to sample a real working day, how to combine the pieces, and how
to attach the normative uncertainty every occupational-hygiene report needs. The
occupational_exposure module adds the three measurement strategies and the
Annex C uncertainty budget on top of the energy-average machinery.
The task-based strategy (Clause 9) splits the nominal day into tasks, takes $I \ge 3$ samples per task, and energy-sums the task contributions
$$ L_{EX,8h,m} = L_{p,A,eqT,m} + 10 \log_{10}(T_m/T_0), \qquad T_0 = 8\ \text{h}, $$
so a loud but short task contributes little. The job-based (Clause 10) and full-day (Clause 11) strategies instead take $N \ge 5$ (or three whole-day) random samples over a homogeneous exposure group and normalise the effective-day duration. The daily level is the same either way; the strategies differ in how the uncertainty is built.
from phonometry.occupational_exposure import (
Task, task_based_exposure, job_based_exposure, full_day_exposure,
)
# ISO 9612 Annex D — a welder's day split into three tasks. Each task level is
# the energy average of its Lp,A,eqT samples; durations carry a measured range.
tasks = [
Task(samples=(70.0,), duration_hours=1.5, label="planning/breaks"),
Task(samples=(80.1, 82.2, 79.6), duration_hours=5.0,
duration_range=(4.0, 6.0), label="welding"),
Task(samples=(86.5, 92.4, 89.3, 93.2, 87.8, 86.2), duration_hours=1.5,
duration_range=(1.0, 2.0), label="cutting/grinding"),
]
res = task_based_exposure(tasks, include_duration_uncertainty=False, warn=False)
print(f"LEX,8h = {res.lex_8h:.1f} dB U = {res.expanded_uncertainty:.1f} dB")
# LEX,8h = 84.3 dB U = 2.7 dB
print(f"one-sided 95 % upper limit LEX,8h + U = {res.upper_limit:.1f} dB") # 87.0 dB
for t in res.tasks:
print(f" {t.label:<16} Lp,A,eqT = {t.lp_aeqt:5.1f} contributes {t.lex_8h_contribution:5.1f} dB")
# planning/breaks Lp,A,eqT = 70.0 contributes 62.7 dB
# welding Lp,A,eqT = 80.8 contributes 78.7 dB
# cutting/grinding Lp,A,eqT = 90.1 contributes 82.8 dB
# The same shift measured job-based (Annex E) and full-day (Annex F): both use
# the Eq C.9 / Table C.4 sampling budget with k = 1.65 (one-sided 95 %).
job = job_based_exposure([88.1, 86.1, 89.7, 86.5, 91.1, 86.7], effective_duration_hours=7.5)
full = full_day_exposure([88.0, 91.9, 87.6, 90.4, 89.0, 88.4], effective_duration_hours=9.25)
print(f"job LEX,8h = {job.lex_8h:.1f} dB U = {job.expanded_uncertainty:.1f} dB")
# job LEX,8h = 88.2 dB U = 3.8 dB
print(f"full-day LEX,8h = {full.lex_8h:.1f} dB U = {full.expanded_uncertainty:.1f} dB")
# full-day LEX,8h = 90.1 dB U = 3.4 dB
Two subtleties are worth spelling out. First, the coverage factor is
$k = 1.65$ for a one-sided 95 % interval (Clause 14), because a hygienist
cares only about the upper bound: res.upper_limit = $L_{EX,8h} + U$ is the
value 95 % of measurements fall below, the number compared against an action
limit. Second, the task and job methods weight the same spread of samples
differently. The task sampling uncertainty $u_{1a}$ (Eq. C.6) divides the summed
squared deviations by $I(I-1)$ — the standard error of the mean, smaller by a
factor $\sqrt{I}$ — whereas the job/full-day sampling uncertainty $u_1$ (Eq. C.12)
is the plain sample standard deviation with denominator $N-1$, whose contribution
$c_1 u_1$ is then read from Table C.4 as a function of $(N, u_1)$. The same
raw scatter therefore inflates the job estimate more, which is the standard's
built-in penalty for coarser, fewer samples. (The printed job $L_{EX,8h}$ is
$88.2$ dB where Annex E reports $88.1$: the standard rounds the effective-day
level to $88.4$ before the duration normalisation; the library keeps it
unrounded.)
When a task's samples span 3 dB or more (Clause 9.3), or the job contribution
$c_1 u_1$ exceeds 3.5 dB (Clause 10.4), or too few workers are covered
(Table 1 cumulative-duration), the result sets sampling_advisory=True and, with
warn=True, emits an ExposureWarning recommending more measurements. Peak
levels $L_{p,Cpeak}$ are reported without an uncertainty — Annex C gives no
method for them (Table C.5, Note 1), so peak-uncertainty is out of scope. The
three Annex D/E/F worked examples above are reproduced to the standard's printed
precision (Annex E's final rounding is disclosed above), and the theory is
derived on the Theory page.
task_based_exposure() / job_based_exposure() / full_day_exposure() parameters #
| Parameter | Applies to | Type | Units | Range / default | Notes |
|---|---|---|---|---|---|
tasks |
task | list of Task |
— | ≥ 1 | Each Task has samples, duration_hours, optional duration_range/duration_samples, label, instrument |
samples |
job / full-day | sequence | dB | ≥ 2 (≥ 5 / ≥ 3 advised) | Random Lp,A,eqT samples |
effective_duration_hours |
job / full-day | float | h | > 0 | Effective working-day duration $T_e$ |
instrument |
all | str | — | 'class1', 'class2', 'personal_exposimeter' (default) |
Selects $u_2$ (Table C.5) |
u3 |
all | float | dB | default 1.0 |
Microphone-position uncertainty (Clause C.6) |
include_duration_uncertainty |
task | bool | — | default True |
False omits the $(c_{1b}u_{1b})^2$ term (Annex D case a) |
n_workers / sample_duration_hours |
job | int / float | — / h | default None |
Table 1 cumulative-duration check |
warn |
all | bool | — | default True |
Emit ExposureWarning for the sampling advisories |
All three return an ExposureResult with lex_8h, combined_standard_uncertainty
$u$, expanded_uncertainty $U = 1.65\ u$, upper_limit = $L_{EX,8h} + U$,
sampling_advisory, and (task-based) the per-task tasks breakdown.
Loudness level of pure tones (ISO 226:2023) #
The normal equal-loudness-level contours relate the SPL of a pure tone to its
perceived loudness level in phons (the SPL of an equally loud 1 kHz tone).
equal_loudness_contour(phon) evaluates ISO 226:2023 Formula (1) at the 29
preferred third-octave frequencies of Table 1, loudness_level(spl, frequency)
is the exact inverse (Formula 2), and hearing_threshold() returns the
threshold-of-hearing column:
from phonometry import equal_loudness_contour, loudness_level
freqs, spl = equal_loudness_contour(40.0) # the classic 40-phon contour
phon = loudness_level(73.0, 63.0) # 73 dB @ 63 Hz -> 40 phon
Validity per clause 4.1: 20-90 phon (80 phon above 4 kHz); the implementation is verified against the Annex B tables in CI. Note this is the loudness of pure tones - loudness of arbitrary signals (ISO 532 sones) is a different, upcoming feature.
Prominent discrete tones (ECMA-418-1) #
Tonal components in machinery noise are far more annoying than their level
suggests. ECMA-418-1:2024 (referenced by ECMA-74 Annex D) gives two FFT-based
methods to decide whether a discrete tone is prominent:
tone_to_noise_ratio() compares the tone level with the masking noise in its
critical band (clause 11), and prominence_ratio() compares the critical band
centred on the tone with the two contiguous bands (clause 12). Both return a
structured verdict against the frequency-dependent prominence criteria:
import numpy as np
from phonometry import tone_to_noise_ratio, prominence_ratio
fs = 48000
rng = np.random.default_rng(0)
t = np.arange(fs) / fs
x = np.sin(2 * np.pi * 1000 * t) + 0.05 * rng.standard_normal(fs) # 1 kHz tone in noise
tnr = tone_to_noise_ratio(x, fs) # highest peak, or tone_freq=...
pr = prominence_ratio(x, fs, tone_freq=1000.0)
print(tnr.ratio_db, tnr.criterion_db, tnr.prominent)
The methods hinge on the critical band — the ear's analysis bandwidth, $\Delta f_c = 25 + 75\ [1 + 1.4(f/1000)^2]^{0.69}$ Hz (162 Hz at 1 kHz): a tone is masked only by the noise inside its critical band, so both ratios compare the tone against exactly that noise, not the whole spectrum.
A TNR above $8 + 8.33\log_{10}(1000/f_t)$ dB (8 dB from 1 kHz up) classifies the tone as prominent; the PR criterion is $9 + 10\log_{10}(1000/f_t)$ dB. Low frequencies get higher thresholds because wider relative bands mask more.
ECMA-74 (which delegates its tone assessments to ECMA-418-1) also fixes where to measure around a device:
Proximate secondary tones in the same critical band are combined per
clause 11.6; for harmonic complexes assess each component (tone_freq=).
Both methods work on Hann-windowed, RMS-averaged spectra and need no absolute
calibration (the ratios are level differences).
tone_to_noise_ratio() / prominence_ratio() parameters #
| Parameter | Type | Units | Range / default | Notes |
|---|---|---|---|---|
x |
1D array | any (uncalibrated OK) | ≥ fs/resolution_hz samples |
Ratios are level differences: calibration cancels out |
fs |
int | Hz | > 0 | |
tone_freq |
float, optional | Hz | 89.1–11 200; default None |
None assesses the highest peak in the range of interest |
resolution_hz |
float | Hz | > 0; default 1.0 |
Tone band must stay within 15 % of the critical band (clause 11.2) |
Both return a ToneAssessment(frequency, ratio_db, criterion_db, prominent).
Environmental noise: Lden, Ldn and rating levels (ISO 1996-1) #
Regulatory noise assessment weights evenings and nights more heavily.
lden() implements the day-evening-night level of ISO 1996-1:2016 (3.6.4:
+5 dB evening, +10 dB night, default 12/4/8 h periods — adjustable, since
countries define them differently), ldn() the day-night variant (3.6.5),
and composite_rating_level() the general whole-day composite of clause 6.5
(Formulae 5-6) for arbitrary periods with source or character adjustments
(Table A.1: e.g. +5 dB regular impulsive, +12 dB highly impulsive, +3 to
+6 dB prominent tones):
from phonometry import lden, composite_rating_level
l = lden(63.2, 58.1, 51.4) # from LAeq per period
r = composite_rating_level([(63.2, 12, 0.0), # day
(58.1, 4, 5.0), # evening (+5)
(51.4, 8, 10.0)]) # night (+10) == lden
lden() / ldn() / composite_rating_level() parameters #
| Function | Key parameters | Notes |
|---|---|---|
lden(lday, levening, lnight, hours=(12, 4, 8)) |
period LAeq values [dB]; hours must sum to 24 |
+5 dB evening, +10 dB night (3.6.4) |
ldn(lday, lnight, hours=(15, 9)) |
+10 dB night (3.6.5) | |
composite_rating_level(periods) |
iterable of (level_db, hours, adjustment_db); hours positive, finite and summing to 24 |
General Formulae (5)-(6); adjustments per Table A.1 |
Where you put the microphone changes the number: ISO 1996-2 fixes the receiver positions and their façade corrections:
Combine with laeq() per time period to go from recordings to Lden, and with
tone_to_noise_ratio() / prominence_ratio() to justify tonal adjustments.
Octave Spectrogram (levels over time) #
Short-time fractional-octave analysis: one level per band per window, time-aligned across bands.
from phonometry import OctaveFilterBank
bank = OctaveFilterBank(fs=48000, fraction=3)
levels, freq, times = bank.spectrogram(recording, window_time=0.125, overlap=0.5)
# levels: (bands, frames) — ready for pcolormesh(times, freq, levels)
A logarithmic sweep plus two tone bursts, resolved in time and in standardized 1/3-octave bands.
- Multichannel input
(channels, samples)returns(channels, bands, frames). timesholds each window's center in seconds.mode='peak'gives per-window peak-holding levels instead of RMS.zero_phase=Truefilters bands forward-backward so per-band group delay does not skew the frames (offline analysis only).
OctaveFilterBank.spectrogram() parameters #
| Parameter | Type | Units | Range / default | Notes |
|---|---|---|---|---|
x |
1D or 2D array | digital units | non-empty | 2D returns (channels, bands, frames) |
window_time |
float | s | > 0; default 0.125 |
Frame length (0.125 s mirrors Fast) |
overlap |
float | — | 0 ≤ overlap < 1; default 0.5 |
Fraction of window overlap (0 = none) |
mode |
str | — | 'rms' (default) or 'peak' |
Per-window detector |
detrend |
bool | — | default True |
Remove each band's DC offset before the level (improves low-frequency accuracy) |
zero_phase |
bool | — | default False |
Forward-backward filtering (offline only) |
calibration_factor / dbfs |
— | — | constructor-only | Set on OctaveFilterBank(...), not per call |
import matplotlib.pyplot as plt
fig, ax = plt.subplots()
mesh = ax.pcolormesh(times, freq, levels, shading="auto")
ax.set_yscale("log")
ax.set_xlabel("Time [s]")
ax.set_ylabel("Frequency [Hz]")
fig.colorbar(mesh, label="Level [dB]")
See Calibration and dBFS to convert digital units to physical SPL, and Time Weighting for the envelope details.