---
Title: AudioLab – measurement methods
Version: 0.2.0
---

# {{ Title }}

How AudioLab computes each result: the signals it plays, the formulas it applies, and the
assumptions those formulas rest on. The formulas are the ones AudioLab actually calculates with;
where a choice differs from a standard, it says so.

## Overview: the mathematics of each task

Nearly everything is done in the **frequency domain**: a block is captured, windowed and transformed by an
FFT, and levels are read as the power of the bins that belong to a tone or a band. AudioLab does not filter the
captured signal in time to measure it – the one exception is the demodulation of a tape tone. What an analog
analyser does with a filter (a notch, a weighting network) is done by leaving out or weighting bins.

| Task | Mathematics | Filtering of the signal |
| --- | --- | --- |
| [Spectrum](#4-spectrum), [Level meter](#level-meter) | windowed FFT (Blackman-Harris, Hann), power averaging, the window's noise bandwidth; the meter's RMS and peak from the samples | none |
| [THD+N, THD, SINAD](#5-thd-thdn-and-noise) | one FFT; a tone's level = the power of its band over the noise bandwidth | **notch = bins left out**: the fundamental ±1 % is not summed; no filter |
| Scope (residual), [section 5](#5-thd-thdn-and-noise) | **least-squares fit** of the fundamental ($a \cos + b \sin + c$), subtracted | the time-domain counterpart of a notch that follows its tone |
| [Noise, S/N, weightings](#6-noise-and-signal-to-noise-ratio), [noise density](#noise-density) | band power of the spectrum, per 1 Hz for density | A, ITU-R 468 and CCIR weightings **multiply the bins** by the curve – no weighting filter |
| [IMD](#7-intermodulation-imd), [Multitone](#multitone) | FFT, tone and product levels from their bins; the multitone is built by an **inverse FFT** (tones on odd bins) | bins left out |
| [Gain & phase](#stimulus-periodic-chirp), Tape response | periodic chirp built by an **inverse FFT**; transfer function **$H_1 = \sum Y X^{*} / \sum X X^{*}$** (cross spectrum, [coherence](#transfer-function-coherence)) | division in the frequency domain – a deconvolution, no convolution in time |
| [Delay](#delay) of Gain & phase | impulse response = inverse FFT of $H_1$ (a cross-correlation), its peak; then the **real cepstrum** gives the minimum phase, so the filters' group delay stays in the phase and only the transport delay is removed | none |
| [Stepped sine](#stepped-sine) | **DFT of a single bin** (Hann window) of output and reference, their ratio; delay from the slope of the unwrapped phase | none |
| [Inverse RIAA](#deviation-from-riaa) on the generator | the periodic stimulus: FFT → each bin divided by the RIAA curve → inverse FFT; a sine: its level scaled | filtering in the frequency domain |
| [Device delay](#device-delay) | RMS envelope in 10 ms windows, the onset of the tone | none |
| [Crosstalk](#9-crosstalk), [Output Z](#output-impedance), [Player](#player-test-file), [Test tape](#test-tape-a-sequence-of-tones) | tone levels from their bins and their ratios; the J-test turns sidebands into jitter by phase modulation | none |
| [Speed & flutter](#10-tape-recorder-speed-wow-flutter-azimuth) | **complex demodulation**: the tone mixed to 0 Hz, **4th-order Butterworth low-pass (IIR)**, decimation to about 1 kHz; instantaneous frequency from the phase step; bands and the speed spectrum by FFT; azimuth from the phase between L and R | the only time-domain filter of the analysis; the IEC 60386 flutter weighting is applied without phase, in the frequency domain |
| [Test record](#11-turntable-the-test-record) | the steps, sweeps and ramps of the tone trackers put to the record's tracks by frequency, cut (level and L/R phase) and order; crosstalk and azimuth averaged over a play; tracking from THD per track | none |

The simulator's devices (filters, RIAA preamplifier, tape recorder) do use IIR filters, but only to make the
simulated signal; the measurement treats it like any other.

## 1. Conventions

| Symbol | Meaning |
| --- | --- |
| $f_s$ | sample rate |
| $N$ | FFT length (block length), a power of two |
| $\Delta f = f_s / N$ | bin width |
| $k$ | bin index, frequency $f_k = k \, \Delta f$ |
| $x[n]$, $y[n]$ | reference and device-output samples of one channel |
| $X_k$, $Y_k$ | their discrete Fourier transforms |
| $w[n]$ | analysis window |
| $L$ | a level in dBFS (section below); $L_{\text{gen}}$ is the generator level set in SIGNAL → *Level* |
| $M$ | number of blocks (or periods) averaged – ADVANCED → *Averages* / *Averages per step* |
| $^{*}$ | complex conjugate ($X^{*}$ has the opposite phase of $X$) |
| $\arg$ | phase (angle) of a complex number |

### dBFS are sine-referenced

A full-scale sine has peak 1.0 and reads **0 dBFS**. Every level in AudioLab, of a tone or of
noise, is expressed as the level of the sine with the same power:

$$
L = 10 \log_{10} P, \;\;\; P = \text{(peak amplitude of the equivalent sine)}^2
$$

A sine of peak $A$ has $P = A^2$; noise of RMS value $\sigma$ has $P = 2\sigma^2$. Levels of tones
and of noise can therefore be subtracted directly (S/N, THD+N), and one calibration offset converts
both to volts (section 11).

### No test signal, clipping

- **No test signal:** a tone (the fundamental for THD+N and the sweeps, the higher tone for IMD, the
  driven channel for crosstalk) below **−100 dBFS** is not evaluated: the result reads *No test signal*
  instead of a ratio of noise to noise. In crosstalk such a point gets no value and is left out of the curve.
- **Clipping:** the level meter marks an input whose largest sample reaches **−0.1 dBFS**; the live
  tasks report *Input clipping detected* when any sample of an analysed block reaches 0.999 of full
  scale. Results taken while clipping are not valid.

## 2. The measurement chain

```mermaid {#fig:chain}
flowchart LR
    GEN["Generator (interface OUT)"] --> DUT["Device under test"]
    GEN --> REF["Reference input"]
    DUT --> OUT["Output input"]
    REF --> CAP["One capture buffer, all inputs sample-aligned"]
    OUT --> CAP
    CAP --> ANA["Analysis per channel pair"]
```

The generator and all inputs run on the **same converter clock** (one ASIO duplex stream), and
every input is written into one ring buffer frame by frame. A block read for analysis therefore
contains the reference and the output of both stereo channels from the same instant. Everything
below rests on this:

- **No clock drift** between what is played and what is recorded, so a tone tuned to a bin centre
  stays on it (section 3).
- **The reference is the device input.** It is taken at the device's input terminal, so
  reference-versus-output comparisons measure the device alone. Loading of the generator by the
  device is included in the reference, as it should be.
- **The reference shows the floor.** Distortion and noise of the generator and converters appear in
  the reference; a device result near the reference is limited by the setup.

A channel pair is (reference input, output input). A stereo wiring has two pairs, L and R, measured
from the same capture.

### External DAC: a second clock and no reference

A headphone amplifier with a USB DAC is measured with the signal played through the DAC (Windows
audio, exclusive mode, so the samples reach the DAC unchanged) while the interface only records:

```mermaid {#fig:dac}
flowchart LR
    GEN["Generator"] --> DAC["DAC (own clock)"] --> DUT["Amplifier"] --> OUT["Output inputs"] --> CAP["Capture (interface clock)"]
```

Two of the assumptions above then fail, and AudioLab uses only methods that do without them:

- **Two clocks.** A tone set to $f$ comes out at $f\,(1 + \varepsilon)$, where $\varepsilon$ is the
  clock difference, typically tens of ppm. At 1 kHz and 30 ppm that is 0.03 Hz, about 1 % of a bin
  at $N = 16384$. The tone is found by peak search and its power summed over the main lobe, so its
  level is unaffected; but even 1 % of a bin makes a Hann window leak enough to put the THD+N floor near
  −70 dB. With a second clock AudioLab therefore always analyses with the Blackman-Harris window, whose
  far sidelobes keep the floor near −110 dB (simulator: −72.7 dB with Hann, −110.6 dB with
  Blackman-Harris at 30 ppm). A periodic stimulus no longer fits the block exactly, so the chirp
  response is not available.
- **No reference.** No input sees the played signal. Gain is taken against the generator's digital
  level (dBFS at the DAC), reference curves and the interface floor are not shown, and reference
  subtraction is off. The pair's "reference" is its own output input, $L_{\text{ref}} = L_{\text{out}}$.
- **Latency.** Windows playback adds 100–200 ms, so every stepped measurement waits 300 ms longer.

## 3. Test signals

### Bin-centred tones

**What a bin is.** The FFT splits a block of $N$ samples into frequency slots called **bins**, all of the
same width

$$
\Delta f = \frac{f_s}{N}
$$

At $f_s = 48$ kHz and $N = 16\,384$ a bin is $48\,000 / 16\,384 = 2.93$ Hz wide; bin $k$ is centred on $k \, \Delta f$.

**Why it matters.** A sine whose frequency sits exactly on a bin centre completes a **whole number of
cycles** in the block. Its energy then lands in that one bin only. A sine between two bins does not fit the
block: the block cuts it off mid-cycle, and part of its energy **leaks** into the neighbouring bins. The
leak can be much larger than the harmonics and noise being measured; a window reduces it, but never removes it.

**What AudioLab does.** The frequency you type in SIGNAL – call it $f_{\text{set}}$ – is moved to the nearest
bin centre $f_{\text{gen}}$, and the generator plays that:

$$
f_{\text{gen}} = \text{round}\left(\frac{f_{\text{set}}}{\Delta f}\right) \Delta f
$$

Step by step for $f_{\text{set}} = 1000$ Hz, $f_s = 48$ kHz, $N = 16\,384$:

1. bin width $\Delta f = 48\,000 / 16\,384 = 2.9297$ Hz,
2. $1000 / 2.9297 = 341.33$ bins – between two bins,
3. rounded to bin 341,
4. $f_{\text{gen}} = 341 \times 2.9297 = 999.02$ Hz.

The generator plays **999.02 Hz** instead of 1000 Hz, and the program shows that frequency. Its harmonics
($2 f_{\text{gen}}$, $3 f_{\text{gen}}$ …) fall on bins 682, 1023 … – also exact bin centres – so they too stay in their own bins.
This is what makes THD readings near −110 dB possible. The offset is never more than half a bin
(1.5 Hz here, 0.37 Hz with $N = 65\,536$).

**997 Hz instead of 1 kHz.** AES17, the standard for measuring digital audio equipment, uses 997 Hz. Its
harmonics (1994, 2991, 3988 Hz …) lie a few hertz beside the whole kilohertz, where interference with a 1 ms
rhythm (USB frames) and the harmonics of 50 Hz mains hum sit; at 1 kHz they fall on them and count as
distortion. And at 48 kHz one period of 1 kHz is exactly 48 samples, repeated over and over, so the
converter's rounding error repeats too and appears as harmonics; 997 Hz never repeats the same samples, and
the rounding spreads into noise. For converters, USB cards and audio interfaces 997 Hz is the tone to use
(the key *997 Hz*; the generator still plays the nearest bin centre, 996.83 Hz with $N = 65\,536$).

The tuning depends on the sample rate and the FFT length: changing either retunes the generator (with
$N = 65\,536$ at 48 kHz, 1 kHz becomes 999.76 Hz). The player's test file has its tones on
the bins of the FFT length it was created for (so keep that length while measuring); a test tape cannot be
tuned at all, and its tones are found wherever the tape speed puts them (sections 8 and 10).

**Generator level.** The level you set in SIGNAL → *Level*, $L_{\text{gen}}$ in dBFS, is the **peak** of the sine
relative to full scale: the sine has amplitude $A = 10^{L_{\text{gen}}/20}$. For example −6 dBFS gives
$A = 0.50$, −20 dBFS gives $A = 0.10$ of full scale.

### Periodic stimuli

The frequency-response chirp (section 8) and the IMD two-tone signal (section 7) are stored as one
period of exactly $N$ samples and repeated. Any block of $N$ captured samples then holds exactly one
period, whatever the delay through the device, so no window is needed and no leakage occurs.

### Stepped measurements

Sweeps, noise and crosstalk change the generator in steps. After each change AudioLab waits for the
settling time and then captures the averaged blocks, reading only frames recorded after the change:

```mermaid {#fig:steps}
flowchart LR
    A["Set generator"] --> B["Wait: settle time"]
    B --> C["Capture M blocks of N samples"]
    C --> D["FFT each block, power-average"]
    D --> E["Analyse every channel pair"]
    E --> A
```

$$
T_{\text{step}} = T_{\text{settle}} + \frac{M \, N}{f_s}
$$

where $T_{\text{step}}$ is the time one step takes, $T_{\text{settle}}$ the settling time (ADVANCED → *Settle*),
$M$ the averages per step and $N / f_s$ the length of one block (1.37 s for $N = 65\,536$ at 48 kHz).

The settling time must exceed the device's own settling. The default is 200 ms; with an external DAC 300 ms
are added for its playback latency.

### Device delay

A tape recorder plays back what it recorded a head spacing later – tenths of a second and more at slow speeds –
so a step would otherwise read the previous step. Before the first step of a sweep (THD, IMD, stepped sine,
crosstalk, and noise with a test tone) AudioLab therefore measures the device's delay:

1. It plays 1 kHz at −20 dBFS (never louder than the loudest step) and records the output, including 100 ms of
   silence before the probe.
2. The envelope is the RMS in 10 ms windows; the level is the highest median of three neighbouring windows.
3. The onset is the first window from which three windows in a row lie within 6 dB of that level, with at least
   12 dB of contrast to the silence and 0.2 s of signal after it; at most 3 s are searched.

Every step then waits $T_{\text{settle}} + T_{\text{delay}}$, and later RUN passes reuse the result. An envelope
ignores wow, flutter and level changes of the tape, which a phase or correlation method would not. SINGLE in the
live tasks waits for the same onset in its capture before it analyses.

## 4. Spectrum

### Window and amplitude scale

A block is multiplied by a window $w[n]$ and transformed. A window lowers the reading by its
**coherent gain** – its average value $\frac{1}{N}\sum w[n]$ (table below) – so the result is divided by
$\sum w[n]$ instead of $N$; then a bin-centred sine of peak $A$ reads $A$ in its bin. The factor $c_k = 2$
folds the negative frequencies onto the positive ones (a real signal has half its power in each); DC and
the Nyquist bin have no mirror image:

$$
A_k = \frac{c_k \, \left| \sum_{n=0}^{N-1} w[n] \, x[n] \, e^{-j 2\pi k n / N} \right|}{\sum_{n=0}^{N-1} w[n]}, \;\;\; c_k = \begin{cases}
1 & k = 0 \text{ or } k = N/2 \\
2 & \text{otherwise}
\end{cases}
$$

### Noise bandwidth and band power

A window spreads noise over more than one bin. Its equivalent noise bandwidth is

$$
B = \frac{N \sum w[n]^2}{\left(\sum w[n]\right)^2} \text{ bins}, \;\;\; \text{ENBW} = B \, \Delta f
$$

and the power in bins $k_1 \ldots k_2$ is

$$
P(k_1, k_2) = \frac{1}{B} \sum_{k=k_1}^{k_2} A_k^2
$$

This one formula gives the power of a tone (summed over its main lobe) and of noise (summed over a
band) on the same sine-referenced scale.

| Window | Coherent gain | ENBW [bins] | Main lobe half-width used [bins] |
| --- | --- | --- | --- |
| Rectangular | 1.000 | 1.00 | 1 |
| Hann | 0.500 | 1.50 | 2 |
| Blackman-Harris (4-term) | 0.359 | 2.00 | 4 |
| Flat top | 0.216 | 3.77 | 5 |

Tone and harmonic powers are summed over the main lobe plus one bin on each side, so that tones a
little off a bin centre are still captured completely.

### Averaging

Averaging is a power (RMS) average of $M$ spectra:

$$
\bar{A}_k = \sqrt{\frac{1}{M} \sum_{m=1}^{M} A_{k,m}^2}
$$

It steadies noise readings; it does not lower the noise floor (that would need synchronous
time-domain averaging). In the live tasks consecutive captures overlap, so $M$ averages steady the
noise less than $M$ independent blocks would.

### Noise density

The noise floor of a spectrum depends on the FFT length and window: a bin holds the noise of one ENBW. With
*Noise spectra as density* (Settings → Chart) the spectra of Spectrum and Noise are shifted by
$-10 \log_{10}(\text{ENBW in Hz})$ and read in dBFS/√Hz (V/√Hz after calibration), the same at any FFT length.
The noise details give the density at 1 kHz as the power of the third-octave band around it divided by its width
in Hz.

### Peak markers

*Peaks* marks the ten strongest lines of each output. A candidate is the highest bin within half the window's main
lobe on either side. It must lie above the local floor – the median of the bins beside it (48 on each side, beyond
twice the lobe) – by a margin from the spread of the noise over the band, $3\,(L_{90} - L_{50})$ and at least 6 dB,
with $L_{90}$ and $L_{50}$ the 90th percentile and median of the band's levels. One capture of white noise has
$L_{90} - L_{50} \approx 5.1$ dB, so the margin is 15 dB (its highest bin lies some 10 dB above the median);
averaging narrows the spread and weaker lines are marked. The harmonics of the strongest line ($k f_0$,
$k = 2 … 10$, within the lobe) take $2\,(L_{90} - L_{50})$, at least 5 dB. A floor rising towards DC is local and
does not pass for lines. Candidates are taken strongest first;
one under the sidelobe envelope of a stronger line (plus 6 dB) is leakage of the window, not a line of its own,
and is skipped. The envelope starts at the first sidelobe (Blackman-Harris −92 dB, Hann −31.5 dB, rectangular
−13.3 dB) and falls 6 dB per octave (Hann 18). Frequency and level come from a parabola through the three bins in dB:

$$
\delta = \frac{1}{2}\,\frac{L_{k-1} - L_{k+1}}{L_{k-1} - 2L_k + L_{k+1}}, \qquad
f = (k + \delta)\,\Delta f, \qquad L = L_k - \tfrac{1}{4}(L_{k-1} - L_{k+1})\,\delta
$$

### Spectrogram

Each capture (not the average) is reduced to 480 linear columns from 0 Hz to Nyquist – the strongest bin
in each column – and added as a row; the last 300 rows (about a minute) are shown as a heat map. The colour
range follows the fixed level axis when it is set, else the 5th percentile to the maximum. The frequency axis
and the plot margins are the same as the spectrum's, so a line stays in place when the view is switched.

### Level meter

For each input the meter reports the true RMS without DC on the sine-referenced scale, the strongest
tone and the largest sample:

$$
L_{\text{RMS}} = 20 \log_{10}\left(\sqrt{2} \sqrt{\frac{1}{N}\sum_{n} (x[n] - \bar{x})^2}\right), \;\;\;
L_{\text{peak}} = 20 \log_{10} \max_n |x[n]|
$$

Here $\bar{x}$ is the mean of the block (its DC component, removed before the RMS). The tone level is $P_1$ of section 5 from a Blackman-Harris FFT. A sine reads the same
$L_{\text{RMS}}$, tone level and (without offset) peak; noise of RMS $\sigma$ reads
$20\log_{10}(\sigma\sqrt{2})$. With a calibration (section 11) the RMS is also given as
$V_{\text{RMS}}$, dBV and dBu, where $\text{dBu} = \text{dBV} + 20\log_{10}(1/0.7746) = \text{dBV} + 2.218$.

## 5. THD, THD+N and noise

With the fundamental (the test tone, frequency $f_1$) at bin $k_1$, its $h$-th harmonic at bin $k_h$, and the
main-lobe half-width $\ell$ (window table plus one bin):

$$
P_1 = P(k_1 - \ell, \, k_1 + \ell), \;\;\; P_h = P(k_h - \ell, \, k_h + \ell), \;\; k_h \approx h \, k_1, \; h = 2 \ldots H
$$

$P_1$ is the power of the fundamental, $P_h$ of harmonic $h$. Harmonics are included up to the chosen order $H$
(ANALYSIS → *Up to harmonic*) and up to the band edge $f_{\text{upper}}$ (ANALYSIS → *Band up to*).

$$
\text{THD} = \sqrt{\frac{\sum_{h=2}^{H} P_h}{P_1}}, \;\;\;
\text{THD+N} = \sqrt{\frac{P_{\text{res}}}{P_1}}
$$

$P_{\text{res}}$ is the power of every bin in 20 Hz – $f_{\text{upper}}$ except the fundamental's band:
harmonics, noise and hum. **Noise** is the same sum with the harmonic lobes also removed, reported
in dBFS. Percent and dB are two views of the same ratio: $\text{dB} = 20 \log_{10}(\text{ratio})$.

- **Notch ±1 %.** The fundamental's band is ±1 % of its frequency, at least the window's main lobe, and $P_1$
  is the power of the whole band – as the notch of an analog analyser takes it. A tone smeared by wow and
  flutter (a tape recorder or turntable) then keeps its sidebands: with ±0.05 % wow and ±0.03 % flutter a main-lobe
  notch read 10 % THD+N at 1 kHz and 250 % at 10 kHz, the ±1 % notch 0.7 %. Harmonics are summed over ±1 % only
  when the tone is smeared (its band holds more than 0.1 dB above its lobe), otherwise over the main lobe, so
  noise beside them does not add to THD.
- **SINAD** is the same measurement the other way up, signal + noise + distortion over noise + distortion:
  $\text{SINAD} = 10 \log_{10}(1 + 1/\text{THD+N}^2)$, so −100 dB THD+N is 100 dB SINAD. Settings → Analysis →
  *Distortion figure* shows it instead of THD+N throughout; the CSV carries both.

- **Largest spur.** The strongest local peak in the band outside the fundamental's lobe (widened by one more
  lobe, so an off-centre tone's skirt is not taken for a spur) and outside the harmonic lobes, with its level
  summed over the lobe like a harmonic. It counts in THD+N but not in THD: when it is higher than the
  harmonics, interference (hum, USB, converter tones) and not non-linearity sets THD+N.
- **Resolution limit.** If $f_1 / \Delta f < 2\ell + 1$, neighbouring harmonic lobes overlap and THD
  is not reported ("longer FFT needed").
- **Weighting.** With A-weighting selected, each bin of $P_{\text{res}}$ and noise is multiplied by the
  A-weighting gain (section 6); the fundamental and THD stay unweighted.
- **Scope view.** The fundamental is fitted by least squares ($a \cos + b \sin + c$) over the last 16 periods
  (at least 4096 samples), its frequency refined from the phase change between the two halves; the residual is
  the signal minus the fit, shown in % of the fundamental's peak over four periods from a rising zero crossing.
  Its shape tells the kind of distortion: spikes at the peaks for clipping, steps at the zero crossings for
  crossover distortion.
- **Reference subtraction** (sweeps, optional): an estimate of the device's own THD,
  $\sqrt{\text{THD}_{\text{out}}^2 - \text{THD}_{\text{ref}}^2}$, shown only where the reference THD lies
  at least **6 dB** below the output THD. The formula treats the two sets of harmonics as uncorrelated, like
  noise. They are not: each harmonic has a fixed phase, so the interface's and the device's harmonics add as
  vectors and can reinforce or cancel. The device's own THD therefore lies anywhere between
  $|\text{THD}_{\text{out}} - \text{THD}_{\text{ref}}|$ and $\text{THD}_{\text{out}} + \text{THD}_{\text{ref}}$,
  and the estimate sits in between: within ±3.5 dB with the reference 6 dB down, ±2.5 dB at 10 dB, ±0.9 dB at
  20 dB. Closer than 6 dB the device is no worse than the interface can show, and no estimate is drawn.
- **Sweep axes.** A THD sweep can be drawn over the generator level, the output level or the output power, and
  as distortion in dB, in % on a logarithmic axis (log10 of the percentage), as the single harmonics H2–H5 in dBc
  or as the output level. Only the drawing changes; every quantity comes from the same points.

## 6. Noise and signal-to-noise ratio

The noise of an output is its band power between 20 Hz and $f_{\text{upper}}$, leaving out DC and
the window's main lobe around it:

$$
N_{\text{dBFS}} = 10 \log_{10} P(k_{\text{low}}, k_{\text{high}}), \;\;\;
N_{\text{A,dBFS}} = 10 \log_{10} \frac{1}{B} \sum_{k} \left(G_A(f_k) \, A_k\right)^2
$$

$k_{\text{low}}$ and $k_{\text{high}}$ are the bins of 20 Hz and $f_{\text{upper}}$; $G_A(f)$ is the A-weighting gain at
frequency $f$. A-weighting follows IEC 61672-1, normalised to 0 dB at 1 kHz:

$$
G_A(f) = 10^{2/20} \cdot \frac{12194^2 f^4}{(f^2 + 20.6^2) \sqrt{(f^2 + 107.7^2)(f^2 + 737.9^2)} \, (f^2 + 12194^2)}
$$

The **ITU-R BS.468-4** (formerly CCIR 468) weighting $G_{468}$ uses the standard's analytic response of its reference
network, shifted by +18.2 dB to 0 dB at 1 kHz (+12.2 dB at 6.3 kHz):

$$
h_1 = -4.7373 \cdot 10^{-24} f^6 + 2.0438 \cdot 10^{-15} f^4 - 1.3639 \cdot 10^{-7} f^2 + 1, \;\;
h_2 = 1.3066 \cdot 10^{-19} f^5 - 2.1182 \cdot 10^{-11} f^3 + 5.5595 \cdot 10^{-4} f
$$

$$
G_{468}(f) = 10^{18.2/20} \cdot \frac{1.2463 \cdot 10^{-4} f}{\sqrt{h_1^2 + h_2^2}}
$$

It is applied like $G_A$, to the band power, so the reading is an **RMS** value (as tape recorders quote "weighted,
CCIR", DIN 45405 *Geräuschspannung*); the quasi-peak detector of the full standard is not modelled.

The measurement has two steps with the generator on the same outputs:

1. Optional test tone (bin-centred sine): its level $S$ at the device output is measured.
2. The generator falls silent, the device input stays connected (terminated by the silent
   interface output), and the noise of every output and reference input is measured.

$$
\text{S/N} = S - N_{\text{dBFS}}, \;\;\; \text{S/N}_A = S - N_{\text{A,dBFS}}
$$

Without the test tone, $S = 0$ dBFS (full scale). *S/N against the level at the distortion limit* takes $S$
from the last THD vs level sweep: the output level where THD+N reached the chosen limit (2 or 3 % for a tape
recorder's maximum output level). The noise on the reference inputs is the floor of
the interface; when the output noise is less than 3 dB above it, the result is flagged as limited by
the interface.

**Check:** white noise of RMS $\sigma$ reads $20 \log_{10}(\sigma \sqrt{2}) + 10 \log_{10}(\text{band} / (f_s/2))$, where band $= f_{\text{upper}} - 20$ Hz;
the tests confirm this within 0.3 dB.

## 7. Intermodulation (IMD)

Two bin-centred tones $f_1 < f_2$ are summed into one period and scaled so that the **peak of the
sum** equals the generator level.

| Method | $f_1$ | $f_2$ | Amplitude ratio $A_1 : A_2$ |
| --- | --- | --- | --- |
| SMPTE | 60 Hz | 7 kHz | 4 : 1 |
| DIN | 250 Hz | 8 kHz | 4 : 1 |
| CCIF | 19 kHz | 20 kHz | 1 : 1 |

$a(f)$ is the amplitude of the component at frequency $f$ – the square root of its power summed over the
main lobe, $a(f) = \sqrt{P(\text{lobe at } f)}$; $d_2$ and $d_3$ are the second- and third-order products.

**SMPTE and DIN** measure the sidebands of the high tone:

$$
d_2 = \frac{\sqrt{a(f_2 - f_1)^2 + a(f_2 + f_1)^2}}{a(f_2)}, \;\;\;
d_3 = \frac{\sqrt{a(f_2 - 2f_1)^2 + a(f_2 + 2f_1)^2}}{a(f_2)}
$$

$$
\text{IMD} = \frac{\sqrt{\sum_{n=1}^{3} \left( a(f_2 - n f_1)^2 + a(f_2 + n f_1)^2 \right)}}{a(f_2)}
$$

**CCIF** measures the difference-frequency products:

$$
d_2 = \frac{a(f_2 - f_1)}{a(f_1) + a(f_2)}, \;\;\;
d_3 = \frac{a(2f_1 - f_2) + a(2f_2 - f_1)}{a(f_1) + a(f_2)}, \;\;\;
\text{IMD} = \sqrt{d_2^2 + d_3^2}
$$

These follow common analyser practice; they have not been verified against IEC 60268-3. The tests
compare them with the analytic products of a polynomial non-linearity $y = x + c_2 x^2 + c_3 x^3$.

### Multitone

31 tones on the third-octave centres from 20 Hz to 20 kHz play together, as one period of the FFT block. Each
sits on an **odd** bin, at least 8 bins from the next (a shorter FFT leaves out the lowest tones that would
crowd), so every second-order product – sums, differences, second harmonics – lands on an even bin, in a gap and
never on a tone. Newman phases $\varphi_k = \pi k^2 / K$ keep the crest factor near 12.7 dB; the generator level
is the peak of the sum.

The gaps are the bins farther than 3 bins (Blackman-Harris: 6) from every tone. The total distortion and noise
is

$$
\text{TD+N} = \sqrt{\frac{\sum_{\text{gaps, 20 Hz – 20 kHz}} P_k}{\sum_{\text{tones}} P_{\text{tone}}}}
$$

and per band (edges at the geometric mean between neighbouring tones) the gaps of the band against its tone.
The flatness of the response is the spread of the output tones against the reference tones; without a reference,
of the output tones among themselves. The Hann window is used; with an external DAC Blackman-Harris and the wider
guard.

## 8. Frequency and phase response

### Stimulus: periodic chirp

One period of $N$ samples is built in the frequency domain. Every bin $k$ between 10 Hz and
$\min(22 \text{ kHz}, 0.45 f_s)$ gets a pink power $P_k \propto 1/k$ (−3 dB per octave, like a
logarithmic sweep) and a phase that makes each tone's group delay equal to the share of energy
below it, as a fraction $\tau_k$ of the period:

$$
\tau_k = \sum_{j \leq k} \frac{P_j}{\sum_i P_i}, \;\;\;
\varphi_k = -2\pi \sum_{j \leq k} \tau_j
$$

The group delay is the slope of the phase, $\varphi_k - \varphi_{k-1} = -2\pi\tau_k$, so tone $k$ sounds at time
$\tau_k N$ within the period. With $P_k \propto 1/k$, $\tau_k$ grows with $\log k$: the tones sweep through the period
one after another, every octave for the same time, which keeps the crest factor low.
The inverse FFT gives the period, normalised to peak 1.0.

![One period of the chirp: the waveform, the frequency over time, and two zooms](figures/chirp.png)

### Transfer function, coherence

Over $M$ periods (no window, section 3), with $X_{k,m}$ and $Y_{k,m}$ the spectra of reference and output
in period $m$:

$$
S_{xy,k} = \sum_{m} Y_{k,m} X_{k,m}^{*}, \;\;\; S_{xx,k} = \sum_{m} |X_{k,m}|^2, \;\;\; S_{yy,k} = \sum_{m} |Y_{k,m}|^2
$$

$$
H_1(f_k) = \frac{S_{xy,k}}{S_{xx,k}}, \;\;\;
\gamma^2_k = \frac{|S_{xy,k}|^2}{S_{xx,k} \, S_{yy,k}}
$$

$S_{xy}$ is the cross spectrum, $S_{xx}$ and $S_{yy}$ the power spectra of reference and output; $H_1$ is the
transfer function (gain and phase), $\gamma^2$ the coherence.

- $H_1$ averages out noise at the output, because noise is uncorrelated with the stimulus. It is
  biased low by noise at the **reference** input; with a clean generator on the reference this is
  negligible.
- **Coherence** $\gamma^2$ (0…1) tells how much of the output is explained linearly by the input.
  Points with $\gamma^2 < 0.8$ (noise, distortion, time variance) are not drawn.
- **Power-ratio magnitude** $\sqrt{S_{yy} / S_{xx}}$ does not depend on phase, so speed variations of
  a tape recorder do not lower it, but noise adds to it. The Tape response task always uses it.

### Delay

The delay through the device is found in two steps.

1. **Impulse peak.** The inverse FFT of $H_1$ is the impulse response; its peak, refined by a
   parabola through three samples, gives $d_{\text{peak}}$. With a periodic stimulus the delay is
   known only modulo $N / f_s$ (1.37 s at $N = 65536$, $f_s = 48$ kHz).
2. **Excess phase.** The peak also contains the group delay of the device's filters, which must stay
   in the reported phase. The minimum phase implied by the magnitude comes from the real cepstrum:

$$
c[n] = \text{IFFT}\left\{\ln |H(f_k)|\right\}, \;\;\;
\hat{c}[n] = \begin{cases}
c[n] & n = 0, \; N/2 \\
2 c[n] & 0 < n < N/2 \\
0 & n > N/2
\end{cases}, \;\;\;
\text{FFT}\{\hat{c}\} = \ln|H| + j \varphi_{\text{min}}
$$

The excess phase $\text{arg}\left(H \, e^{j 2\pi k d_{\text{peak}} / N}\right) - \varphi_{\text{min}}$ of a
minimum-phase device plus a delay is a straight line; a least-squares fit over coherent bins away
from the band edges gives the remaining delay correction. Filters drop out; a pure delay remains.

A delay of at least **0.5 ms** is treated as transport delay (tape head spacing, converter latency)
and removed from the reported phase. Shorter delays stay in it, so an electronic device keeps its
true phase.

### Smoothing and summary

The response is averaged over fractional-octave bands (6–48 points per octave) after removing the
measured delay, so the phase rotation of a long delay does not cancel inside a band:

$$
H_{\text{band}} = \frac{\sum_{k \in \text{band}} S_{xy,k} \, e^{j 2\pi f_k \tau}}{\sum_{k \in \text{band}} S_{xx,k}}
$$

where $\tau$ is the measured delay (above); the factor $e^{j 2\pi f_k \tau}$ undoes its phase before the bins are summed.

The gain is read at the point nearest 1 kHz; the −3 dB corners are the first points on each side
that fall 3 dB below it, interpolated in log frequency.

### Player test file

A player with its own converter plays a file AudioLab writes. Each part of the file is a tone on its own
bin centre of the analysis FFT (at least three bins from every other part and from mains frequencies up
to the sixth harmonic), so the part being played is known from the frequency of the loudest tone, within
1.2 bins; silence is the absence of tones with broadband noise below −60 dBFS. A part counts once it has
been heard in 3 consecutive blocks (silence 8), which leaves out blocks across two parts, and of these the
loudest block is kept (the median for silence). The parts: 1 kHz at −1 to −90 dBFS, digital silence, 1 kHz
and 10 kHz in one channel only, 31 tones from 20 Hz to 20 kHz at −10 dBFS, and SMPTE and CCIF. The file's
16- and 24-bit samples carry triangular dither of ±1 LSB.

The file ends with the **J-test**: a tone at $f_s/4$ written as the codes $\pm 2^{b-2}$, one LSB lower in every
other half-period of 192 samples, without dither. Jitter of the converter modulates the phase of the tone and
shows as sidebands at $f_s/4 \pm$ odd multiples of $f_s/192$, shifted by the player's clock. The square wave's own
harmonics, $(2/\pi) \cdot \text{LSB} / h$ for $h = 48 \mp m$, are subtracted in power; a sideband counts when it lies
6 dB above the floor and above that harmonic, and the peak jitter is $t_j = 2 \cdot 10^{\text{dBc}/20} / (2\pi f)$.

From these, with $L_{-1}$ the measured level of the −1 dBFS tone, $L_{\text{step}}$ the measured level of a step
and $L_{\text{nominal}}$ its level in the file: the output level for 0 dBFS in the file is $L_{0\,\text{dBFS}} = L_{-1} + 1$ dB,
the linearity error of a step is $L_{\text{step}} - L_{\text{nominal}} - L_{0\,\text{dBFS}}$, and the dynamic range after
AES17 is full scale over the A-weighted noise and distortion beside the −60 dBFS tone ($L_{-60}$ its measured
level):

$$
\text{DR} = L_{0\,\text{dBFS}} - \left(L_{-60} + 20\log_{10} \text{THD+N}_A(-60)\right)
$$

### Test tape (a sequence of tones)

A test tape plays its own tones at its own speed, so no list of frequencies can be relied on. A block
holds a tone when its strongest component is above −70 dBFS and at least 20 dB above everything else
in the band; the tone counts once it has kept its frequency within 2 % and its level within 2 dB for
four blocks (about a second). Each run of such blocks is one step, averaged, at the frequency heard; a
gap or a change of more than 2 % or 3 dB starts a new one. Steps more than 6 dB above the median
level are reference tones; the others form the series, given relative to its step nearest 1 kHz.

### Stepped sine

The chirp needs the reference sample-exact. Without one — an external DAC on its own clock — the response
is measured tone by tone: at each frequency $f_i$ the tone's level is read at its spectral peak (power
sum over the main lobe, as for THD), and

$$
G(f_i) = L_{\text{out}}(f_i) - \begin{cases} L_{\text{ref}}(f_i) & \text{with a reference} \\ L_{\text{gen}} & \text{without} \end{cases}
$$

in dB, where $L_{\text{out}}$ and $L_{\text{ref}}$ are the tone levels at the output and reference inputs and
$L_{\text{gen}}$ the generator level, all in dBFS. The peak search finds the tone wherever the DAC's clock puts it, and each step waits for the
Windows playback latency. The gain near 1 kHz, the −3 dB corners and the spread over 20 Hz – 20 kHz
are read as for the chirp response.

**Phase** (with a reference only): in every block a Hann-weighted DFT at the tone's own bin on the output and the
reference, $H = \sum Y X^{*} / \sum |X|^2$ over the blocks. A point is drawn when the tone stands 20 dB above the
mean of bins 4–9 beside it, so the phase holds deep in a filter's stop band, where the chirp's coherence fails.
The phase is unwrapped across the steps; the delay is $-\text{slope}/360°$ by least squares and removed from 0.5 ms
up, as for the chirp. When more than a quarter of the neighbouring steps jump by more than 120°, the steps are too
sparse for the delay (a tape recorder) and no phase is shown. An external DAC gives no phase.

### Deviation from RIAA

The ideal RIAA playback curve is

$$
H_\text{RIAA}(s) = \frac{1 + sT_2}{(1 + sT_1)(1 + sT_3)}, \qquad T_1 = 3180\ \mu\text{s},\ T_2 = 318\ \mu\text{s},\ T_3 = 75\ \mu\text{s},
$$

where $s = j 2\pi f$ (the complex frequency) and $T_1 \ldots T_3$ are the time constants of the standard; optionally
times the IEC amendment $sT_4 / (1 + sT_4)$ with $T_4 = 7950\ \mu$s. With a curve chosen, the measured response
– gain $G(f)$ in dB and phase $\varphi(f)$, from the chirp or the stepped sine – is reported as

$$
D(f) = G(f) - 20\log_{10}\frac{|H_\text{RIAA}(f)|}{|H_\text{RIAA}(1\ \text{kHz})|}, \qquad
\varphi_D(f) = \varphi(f) - \arg H_\text{RIAA}(f),
$$

so a perfect preamplifier gives its 1 kHz gain at every frequency and zero phase. The figure of merit is
the largest $|D(f) - D(1\ \text{kHz})|$ over 20 Hz – 20 kHz, reported with its sign and frequency.

With the **inverse RIAA** on the generator, a sine of level $L_{\text{gen}}$ at $f$ plays at
$L_{\text{gen}} - 20\log_{10}\big(|H_\text{RIAA}(f)| / |H_\text{RIAA}(1\ \text{kHz})|\big)$, and a periodic stimulus of length $N$
has each bin $k$ divided by the normalised $H_\text{RIAA}(k f_s / N)$ before the inverse FFT, so it stays
periodic with its tones on their bins. The resulting peak is capped at the output limit.

### Output impedance

A source of open-circuit voltage $U_0$ and impedance $Z$ gives $U_R = U_0 \, R / (R + Z)$ across a load
$R$, so

$$
Z = R \left( \frac{U_0}{U_R} - 1 \right), \quad \frac{U_0}{U_R} = 10^{(L_0 - L_R)/20}
$$

with $L_0$, $L_R$ the tone levels in dBFS without and with the load. Only the ratio enters, so no
calibration is needed. The damping factor is $R / Z$. The tone levels repeat to about 0.005 dB, so
$Z$ below roughly $0.0006\,R$ is not resolved.

### Channel balance (L − R)

With both channels shown, the responses are compared on the same frequency grid:
$\Delta G = G_L - G_R$ (gains of L and R in dB) over points reliable in both channels, and the phase
difference of the phases $\varphi_L$, $\varphi_R$:

$$
\Delta\varphi(f) = \varphi_L(f) - \varphi_R(f) + 360^\circ \, f \, (\tau_R - \tau_L)
$$

where $\tau_L$, $\tau_R$ are the transport delays removed from each channel. The delay difference is
put back so that, for example, a tape head's azimuth appears in the phase difference; it is also
reported directly.

## 9. Crosstalk

At each frequency the generator drives **one output only**, and the level of the tone is read in
its own main lobe (no peak search) on both outputs:

$$
L_{\text{drv}} = 10 \log_{10} P_{\text{driven}}(k), \;\;\; L_{\text{leak}} = 10 \log_{10} P_{\text{undriven}}(k)
$$

where $k$ is the tone's bin, $L_{\text{drv}}$ the tone level on the driven output and $L_{\text{leak}}$ the level in the
same bins on the other output. At the start of each pass the same bins are measured with the generator
silent, giving the background $L_{\text{bg}}$ (hum, noise). The crosstalk, with the background power removed, is

$$
\text{CT} = 10 \log_{10}\left(10^{L_{\text{leak}}/10} - 10^{L_{\text{bg}}/10}\right) - L_{\text{drv}}
$$

Separation is $-\text{CT}$. Where $L_{\text{leak}} - L_{\text{bg}} < 3$ dB the leak cannot be told from the background;
CT is then only an upper bound ($L_{\text{leak}} - L_{\text{drv}}$), such points are marked and left out of the curve. Reading the
tone's own bins (instead of the wideband level) is what allows crosstalk far below the noise to be
measured; not searching for a peak keeps a nearby 50 Hz hum line from being taken for crosstalk.

## 10. Tape recorder: speed, wow, flutter, azimuth

```mermaid {#fig:tape}
flowchart LR
    T["Played-back tone f0"] --> M["Mix to 0 Hz, low-pass, decimate to about 1 kHz"]
    M --> F["Phase step: speed deviation in %"]
    F --> B["FFT bands: drift, wow, flutter"]
```

The played-back tone ($f_0$ = the generator frequency, or the strongest tone found) is mixed to
0 Hz, low-pass filtered (4th-order Butterworth) at $\min(400 \text{ Hz}, f_0/2)$ and decimated to about $r = 1$ kHz, giving the
complex envelope $z[n]$. The instantaneous frequency $f[n]$ and the speed deviation $\delta[n]$ are

$$
f[n] = f_0 + \frac{r}{2\pi} \text{arg}\left(z[n] \, z^{*}[n-1]\right), \;\;\;
\delta[n] = \left(\frac{f[n]}{\bar{f}} - 1\right) \cdot 100 \%
$$

where $\bar{f}$ is the mean of $f[n]$ over the window. The deviation series is split into bands by FFT; each band's RMS is
$\sqrt{P(\text{band})/2}$ in percent:

| Band | Frequencies | Typical cause |
| --- | --- | --- |
| Drift | below 0.5 Hz | slow speed changes |
| Wow | 0.5 – 6 Hz | capstan, reels |
| Flutter | 6 – 200 Hz | rollers, bearings |
| Wow & flutter | 0.5 – 200 Hz | RMS, 2σ and peak |

The **2σ** value is the "2-sigma" reading of AES6 (the same method as IEC 60386 and DIN 45507): the
deviation that the band-limited speed $\delta_b[n]$ exceeds, up or down, for only 5 % of the time,
that is the 95th percentile of $|\delta_b[n]|$:

$$
\text{fraction of } n \text{ with } |\delta_b[n]| > \delta_{2\sigma} = 5 \%
$$

For a Gaussian deviation this is $1.96\,\sigma$, for a sine $\sin(0.95 \cdot 90^\circ) = 0.997$ times its
peak. Unlike the largest peak, it is not set by a single brief excursion. The analysis window is 10 s
(at least 5 s are required), shorter at high sample rates where the capture buffer holds less.

**Speed error** is $(\bar{f} / f_{\text{nominal}} - 1) \cdot 100 \%$; it is meaningful with a test tape
(record and playback on the same machine cancel it). **Azimuth**, the delay of the right channel,
comes from the phase between the two envelopes $z_L[n]$, $z_R[n]$ of the left and right channel:

$$
\tau_{\text{az}} = -\frac{\text{arg} \sum_n z_R[n] \, z_L^{*}[n]}{2\pi \bar{f}}
$$

It is unambiguous only while $|\tau_{\text{az}}| < 1 / (2 f_0)$. **Level** stability comes from $|z[n]|$.

**Weighted** wow and flutter apply the IEC 60386 / DIN 45507 / AES6 curve (0 dB at 4 Hz, where speed variation
is heard most; −30.7 dB at 0.2 Hz, −3.0 dB at 1 Hz, −8.5 dB at 31.5 Hz, −22.5 dB at 200 Hz, linear in dB over
log frequency between the points) to the deviation up to 200 Hz in the FFT, without its phase. From the weighted
series come the **W.RMS** (as JIS, NAB and Japanese recorders quote it) and the **weighted 2σ** (as DIN and IEC
quote it). They are approximate: the curve is read off the standard's published graph to about ±0.5 dB, and the 2σ
statistics stand in for the quasi-peak meter of DIN 45507. The unweighted RMS, 2σ and peak are given as well.

### Spectrum of the speed deviation

A wider demodulation of its own (4 kHz rate, 1.5 kHz low-pass, up to 2/3 of it, at most 1 kHz) gives the
deviation by frequency from 0.15 Hz, in % RMS. The figures above stay on the band up to 200 Hz; *scrape flutter*
above it, the tape rubbing on the heads, is shown but counted in none. Lines are local maxima above three times the
median and 5 % of the largest, their frequency refined by a parabola through the log amplitudes. A roller of
diameter $d$ at tape speed $v$ turns at

$$
f = \frac{v}{\pi d}
$$

so with the speed and the capstan and pinch roller diameters in Settings → Analysis → *Tape transport*, a line
within 4 % is named after them; a turntable's platter is named at 0.56 Hz (33⅓ rpm) or 0.75 Hz (45 rpm) within 3 %.

## 11. Turntable: the test record

AudioLab plays nothing here: the record plays, and its tracks are recognised as they are heard. Three trackers run on
the same blocks (about 0.35 s each: 16384 samples at 48 kHz, 32768 at 96 kHz): **steady tones** become steps as for a
test tape ([above](#test-tape-a-sequence-of-tones)), **sweeps** are followed from block to block, gliding or stepping a
sixth of an octave, and the **anti-skating ramp** as a tone rising in steps of level.

### Recognising the tracks

Each item heard is put to a track of the chosen record by what it is, not by when it comes:

- **A tone** goes to a tone track within 10 % of its frequency – wide enough for a speed error and a pitch control turned
  while it plays – and of its cut: one channel at least 12 dB above the other is a one-channel track, both alike a lateral
  or vertical one, told apart by the correlation of L and R (in phase lateral, out of phase vertical; $|\rho| \geq 0.3$).
  Tracks of one frequency at different levels (a reference and a −20 dB tone) are told by the level against the loudest
  step of that frequency.
- **Tracks that look alike** (Ortofon 5 and 7, 6 and 8) go by the track heard before: the first of them after it on the
  record, from the first again past the last. A one-channel tone right after one of the same channel stays on its track,
  as does a tone whose frequency moves by more than 0.2 % from the step before (the pitch turned).
- **The tracking tracks** (one frequency, rising level) are found as a run of three or more steps, each louder than the one
  before, and put to the tracks by the spacing of their levels.
- **A sweep** goes to the sweep track of its direction and cut whose range lies closest, in octaves at its two ends.
- A track **clicked** before it plays takes the next item of its kind, whatever it looks like.

A track played again replaces the one before. The Setup view leaves out of its record check the tones that sit on the
mains (50 or 60 Hz and their multiples below 400 Hz, within 1 %) unless the record has a track there, and takes the
record as the wrong one when at least three items heard, and a quarter of all, fit none of its tracks.

### Speed, balance, response

**Speed** is $(\bar{f}/f_{\text{nominal}} - 1) \cdot 100\,\%$ of the speed tone heard last (the 3150 Hz tone where the
record has one, else 1 kHz). **Channel balance** is $L - R$ in dB from the last tone of L alone and the last of R alone.
**Response**: each sweep point relative to the sweep's own point nearest 1 kHz, against what the track is cut to – the RIAA curve
for a sweep cut at constant velocity, flat otherwise – within the record's valid band.

The Setup view checks the phono correction by fitting the deviation $d(f)$ of the sweeps to the RIAA curve $R(f)$ in dB:

$$
k = \frac{\sum d(f)\,R(f)}{\sum R(f)^2}
$$

$k \approx 0$ is the correction in place, $k \approx -1$ a correction missing, $k \approx +1$ one applied twice; it is
judged beyond $|k| = 0.5$, and only where the curve reaches at least 6 dB.

### Crosstalk and azimuth

For a tone cut in one channel the leak into the other is $C = L_{\text{other}} - L_{\text{driven}}$ in dB. The azimuth
reading is the **difference** of the two directions,

$$
\Delta = \overline{C}_{L \to R} - \overline{C}_{R \to L}
$$

each direction **averaged over the latest play** of the record's one-channel 1 kHz tracks (Ortofon 5–8, Analogue
Productions A2 and A3). The latest play is found from the tone heard last back through the record's order, as long as each
was heard just before the next (one to three steps apart); a play of 5–8 after the headshell is turned thus leaves the older
tones out. A single pair is not used: pairs of one record read differently, while the average of a whole play repeats
within a few tenths of a decibel. The Setup view judges $\Delta$ once the play holds all the tracks, within ±2 dB. Without a
tone of one direction, its one-channel sweep at 1 kHz stands in.

### Tracking force and anti-skating

On the tracking tracks each channel's THD is read per track; the **tracking ability** is the largest amplitude up to which
it stays under 3 %. The Setup view judges the force and the anti-skating from the two channels together, once the loudest
track has played:

| Reading | Taken as |
| --- | --- |
| R mistracks first, or THD of R at the loudest track ≥ 2 × L and ≥ 1 % | too little force |
| both channels past 3 % | too little force |
| L mistracks first, or L ≥ 2 × R and ≥ 1 % | anti-skating too high |
| one channel worse on every track: $\exp\big(\,\big|\overline{\ln(\text{THD}_R/\text{THD}_L)}\big|\,\big) \geq 1.4$ over the tracks both hold | anti-skating too low (R worse) or too high (L worse) |
| the worse channel at the loudest track ≥ 2 % | good, close to the limit |

On a record with an **anti-skating ramp** (a 315 Hz tone rising in steps), the THD of each channel is read per step and the
level where it passes 3 % found; the two channels within 1 dB of each other, or neither passing, is right.

### Signal checks

| Check | Rule |
| --- | --- |
| Input overload | a tone or ramp step above −4 dBFS RMS (its peaks within 1 dB of full scale) |
| A channel with no signal | one channel at least 40 dB below the other on everything heard |
| A channel that comes and goes | the same channel's one-channel 1 kHz tones, cut at one level, more than 3 dB apart |

## 12. Calibration in volts

Because dBFS are sine-referenced (section 1), one offset per input converts every level:

$$
L_{\text{dBV}} = L_{\text{dBFS}} + 20 \log_{10} V_{\text{FS}} + A_{\text{ext}}
$$

$V_{\text{FS}}$ is the RMS voltage at the input that reads 0 dBFS, found from a sine of known voltage
$V_{\text{applied}}$ (the multimeter reading) that read $L_{\text{meas}}$ dBFS: $V_{\text{FS}} = V_{\text{applied}} / 10^{L_{\text{meas}}/20}$.
$A_{\text{ext}}$ is the loss of an external divider (for example 20 dB for 10 : 1), added back so that
readings refer to the probed point.

During calibration the multimeter reads the interface output itself (it is connected straight to the
input), so the same reading also calibrates the outputs: $V_{\text{FS,out}} = V_{\text{applied}} / 10^{L_{\text{gen}}/20}$, with
$L_{\text{gen}}$ the calibration tone's level (−10 dBFS). With both inputs of a pair calibrated,
the response magnitude becomes the true voltage gain,
$G = |H| + (\text{offset}_{\text{out}} - \text{offset}_{\text{ref}})$ in dB, where the offset of an input is
$20 \log_{10} V_{\text{FS}}$ of that input.

### Output power

With a resistive load $R$ set on the output (a speaker load of 2–16 Ω, or headphones of 16–600 Ω), the calibrated RMS voltage at the load (the divider added
back) gives the power $P = V_{\text{RMS}}^2 / R$. THD vs level then runs over output power. The rated
power at 1 % THD+N is found between the two steps $a$, $b$ where THD+N first rises through 1 %,
interpolated linearly in log power and log THD+N:

$$
\log P_{1\%} = \log P_a + \frac{\log(0.01 / D_a)}{\log(D_b / D_a)} \left(\log P_b - \log P_a\right)
$$

where $D$ is the THD+N ratio. If the sweep stays below 1 %, or starts above it, no rated power is given.

## 13. How the calculations are checked

Every calculation in this document is checked on a signal whose correct result is known in advance.
AudioLab's built-in simulator produces such signals – with a chosen amount of noise, distortion, hum,
filtering, delay, crosstalk or output impedance – and automated tests compare what AudioLab computes with
what was put in.

This checks the **calculations**, not the accuracy of a particular interface: that depends on the
interface's own noise and distortion, which the reference input shows in every measurement, and on the
multimeter used for calibration (section 13).

| What was put in | Correct result | AudioLab computes |
| --- | --- | --- |
| White noise over 20 Hz – 20 kHz, known RMS value | its level in dBFS (section 6) | the same within 0.3 dB |
| Preamplifier with +10 dB gain and known noise, on an interface with known noise | S/N 101.5 dB | 101.1 – 101.2 dB |
| Amplifier crosstalk of −75 dB at 1 kHz, rising 6 dB per octave | −75.0 dB at 1 kHz, −51.6 dB at 20 kHz | −75.0 dB, −51.6 dB |
| High-pass filter at 100 Hz and low-pass at 10 kHz | corners at 100 Hz and 10 kHz; phase +45° and −90° there | 95 – 105 Hz and 9.5 – 10.5 kHz; +43 to +47°, −93 to −87° |
| Tape recorder with 150 ms head delay and filters | delay 150 ms, filter phase kept (section 8) | 150 ms ±4 µs |
| Tape head tilted so the right channel is 15 µs late | azimuth 15 µs | 14.9 µs |
| Headphone amplifier with 2.2 Ω output impedance, 32 Ω load | 2.20 Ω, damping factor 14.5 | 2.20 Ω, 14.5 |
| Preamplifier with +10 dB gain, stepped sine | +10.0 dB | +9.9 to +10.1 dB |
| USB DAC on its own clock (30 ppm off), stepped sine | flat response | within 0.5 dB |
| Ideal RIAA curve at 20 Hz and 20 kHz relative to 1 kHz; with the IEC amendment at 20 Hz | +19.27 / −19.62 dB; +16.26 dB | the same to 0.01 dB |
| Phono preamplifier (simulated RIAA stage) against the ideal curve | within ±0.25 dB | within ±0.3 dB |
| Second-order low-pass, stepped sine with a reference | phase −16.4° at a fifth of the corner, −90° at the corner | the same |
| Tape recorder with wow ±0.05 % and flutter ±0.03 %, THD+N at 1 and 10 kHz | below 3 %, no spur on the tone | 0.7 % |
| Weighting curve at 4 Hz and 0.2 Hz | 0 dB and −30.7 dB | the same |
| Tape recorder with 150 ms head delay, sweep | each step read after the delay | delay found, steps at full level |

Where a range is given, it is the range the automated tests accept.

## 14. Assumptions and limits

- **Linear, time-invariant device** for the frequency response; coherence shows where this fails.
- **Stationary signals** during a block; averaging reduces variance, not bias.
- **Common clock** for generator and inputs; with separate clocks tones drift off their bins. With an
  external DAC this does not hold, so only tone-level methods are offered (section 2).
- **The reference is taken at the device input**; a long cable between the tap and the device adds
  its own response to the result.
- **Delay** is known only modulo one stimulus period.
- **Reference subtraction** of THD and **IMD** formulas are estimates following common practice,
  not certified against IEC 60268-3.
- **Weighted wow and flutter** are approximate: the curve is read off the standard's graph and the 2σ statistics of
  AES6 stand in for the quasi-peak meter.
- **Results near the reference or the interface floor** describe the setup, not the device.
- **Accuracy of a real measurement** is set by the audio interface and the calibration: its noise and
  distortion (visible in the reference), the flatness of its converters, and the multimeter used to
  calibrate. The checks in section 12 cover the calculations only.
