Measurement Practice & Error Analysis · Manual Chap. 13.5, 13.11, 13.17, 16 & 22.4

The 4 Most Common Pitfalls in FFT Spectrum, Noise, and Distortion Measurements

A spectrum analyzer displays curves that look unambiguous at first glance. In laboratory practice, however, we repeatedly encounter four classic misinterpretations: confusing the visual FFT noise floor with true SNR, blaming 32-bit FFT rounding noise on the device under test, losing 3.5 dB of sine level to window scalloping, and masking amplifier hiss by 10 dB with A-weighting.

Pitfall 1 · Spectral Density vs. Wideband RMS (Manual Chap. 13.5.1 & 16.3)

Why a -150 dB Noise Floor in the Spectrum Actually Means -120 dB SNR

Many users look at a high-resolution FFT plot, see the visual noise floor sitting at -90 dB or -150 dB, and record that number as the Signal-to-Noise Ratio (SNR). This is physically incorrect: an FFT spectrum does not show total energy, but spectral density (energy per frequency bin).

If you quadruple the FFT block size N (e.g., from 4,096 to 16,384 points), the exact same noise energy is distributed across four times as many frequency bins—the visible noise floor drops by 6 dB even though the input signal has not changed at all! To obtain the true wideband RMS noise level, the spectral density must be integrated across all N/2 bins. At 4,096 points (2,048 bins), this FFT process gain is 10 · log10(2048) = 33.1 dB. A visual floor at -156 dB integrates to -122.9 dB true RMS noise.

Interactive Simulator: Visual FFT Noise Floor vs. True Wideband RMS Noise

Visible Noise Line in FFT Plot (Spectral Density)
-153.1 dB
Apparent Floor Drop (FFT Process Gain 10·log10(N/2))
+33.1 dB
True Wideband RMS Level / THD+N (Independent of N!)
-120.0 dB
Pitfall 2 · 32-Bit Float vs. 64-Bit Double FFT (Manual Chap. 16 & 22.4)

Why a 32-Bit Floating-Point FFT Generates Its Own Noise at 8,192 Points

At first glance, a 32-bit floating-point variable (IEEE 754 Single Precision) with over 200 dB of exponent range seems more than sufficient for any 16-bit or 24-bit audio signal. However, a 32-bit float only has a 24-bit mantissa (and 8-bit exponent). In a 4,096- or 8,192-point FFT, every sample passes through 12 or 13 sequential butterfly addition stages.

At each stage, tiny rounding errors in the 24-bit mantissa accumulate. Our bench comparison in the manual (Chap. 22.4) using an ideal 16-bit digital sine file proves the effect: at 1,024 points, both the 32-bit and 64-bit FFT reach the theoretical 96 dB THD+N. Increasing resolution to 8,192 points causes the 32-bit FFT's own rounding noise to degrade the reading to 90 dB THD+N (a 6 dB loss!), whereas our 64-bit double-precision FFT (#3005) with a 144 dB lower math floor stays rock-solid at 97 dB.

Test Setup (1 kHz Ideal Sine) Standard 32-Bit Float FFT WinAudioMLS 64-Bit FFT (#3005)
16-Bit Signal · 1,024-Point FFT (10 Stages) 96.0 dB THD+N (OK) 96.0 dB THD+N (OK)
16-Bit Signal · 8,192-Point FFT (13 Stages) 90.0 dB THD+N (6 dB Error from FFT Noise!) 97.0 dB THD+N (Exact)
24-Bit Signal · 4,096 Points (0.0005% THD / 120 dB SNR) 3rd Harmonic k3 Masked by Math Noise 120.0 dB SNR & 0.0005% THD Clearly Resolved
Pitfall 3 · Window Scalloping Loss & Peak Interpolation (Manual Chap. 13.17.1)

3.5 dB Level Error Between FFT Bins – and How Peak Interpolation Corrects to < 0.01 dB

A Discrete Fourier Transform only calculates levels at discrete frequency bins (f = n · F_s / N). At 44,100 Hz and 1,024 points, 1,378.125 Hz hits an exact bin, whereas 1,000.0 Hz falls between bins. Sweeping a pure sine from 100 Hz to 1,000 Hz causes the displayed peak amplitude on simple FFT analyzers to fluctuate periodically by up to 3.5 dB (Rectangular window) or 1.0 dB (Blackman window).

WinAudioMLS solves this in two ways: first via the Flat-Top window for continuous sweeps, and second via our mathematical Peak Interpolation algorithm, which reconstructs the true peak from adjacent bins. This reduces the residual level error with a Blackman window from 1.0 dB down to less than ±0.01 dB (and with a Rectangular window from 3.5 dB to < 0.1 dB).

Pitfall 4 · CCIR / ITU-R 468 vs. dB(A) in Noise Testing (Manual Chap. 13.11.1)

Why dB(A) Reads 10 dB Lower on White Noise Than the ITU-R 468 Curve

The A-weighting curve dB(A) is derived from the 40-phon Fletcher-Munson equal-loudness contour for pure single sine tones. For broadband random noise (amplifier hiss, microphone self-noise, tape hiss), however, the A-curve does not match human perception. The human ear is acutely sensitive to random hiss between 5 kHz and 6.3 kHz.

To address this, the broadcast and studio industry established the CCIR / ITU-R 468 weighting curve: normalized to 0 dB at 1 kHz, it boosts the 6.3 kHz region by +12.2 dB before rolling off high frequencies. Measuring the exact same white noise source (20 Hz – 20 kHz) yields an A-weighted RMS value that is a full 10 dB lower than the ITU-R 468 reading! That 10 dB cosmetic advantage is why marketing datasheets prefer dB(A). WinAudioMLS lets you switch directly between DIN A, DIN C, CCIR/ITU-R 468, and ITU-R ARM.

Interactive Weighting Comparison: Unweighted (Z) vs. DIN A vs. CCIR / ITU-R 468

Unweighted (Linear / dB(Z))
-80.0 dB
A-Weighting (dB(A))
-82.0 dB(A)
CCIR / ITU-R 468 (dB(468))
-72.0 dB(468)

Matching Software, Ultra-Low-Noise Microphones & Lab Calibration

Measurement Systems for High-Resolution Spectral and Noise Analysis

Configure your measurement chain for distortion and noise analysis—from the dual-channel WinAudioMLS PRO EX with 192 kHz ASIO expansion to our 6.5 dB(A) Microtech Gefell microphone set and in-house self-noise lab calibration:

Dr. Jordan DesignArt. #1721↗

Pro EX

extends to 2 channel, THD, Impedance, Spectrogram

Price:279,00 €net
incl. 19% VAT, excl. shipping:332,01 €
Dr. Jordan DesignArt. #1722↗

LAB

Log-Sweep (chirp), 192kHz sample rate, incl. USB Dongle

Price:479,00 €net
incl. 19% VAT, excl. shipping:570,01 €
Dr. Jordan DesignArt. #3000↗

Plug-in ASIO 192kHz 24bits

Expand your measurement system to 24Bit and 192kHz Samplerate via ASIO

Price:99,00 €net
incl. 19% VAT, excl. shipping:117,81 €
Dr. Jordan DesignArt. #7580↗

Microtech Gefell MTG MMS 214 ultra low noise microphone 6.5dB(A)

Microtech Gefell MTG MMS 214 ultra low noise microphone 6.5dB(A)

Price:3150,00 €net
incl. 19% VAT, excl. shipping:3748,50 €
Dr. Jordan DesignArt. #75240↗

MTG Microtech Gefell MV240 USB Interface 1/2"

MTG Microtech Gefell MV240 USB Interface 1/2"

Price:1963,03 €net
incl. 19% VAT, excl. shipping:2336,00 €
Dr. Jordan DesignArt. #S101↗

microphone calibration self noise

we measure the self noise of the microphone and pre-amplifier

Price:150,00 €net
incl. 19% VAT, excl. shipping:178,50 €
More WinAudioMLS DSP Simulators
Acoustic Self-Noise
Low-Noise Measurement Microphones (< 15 dB(A) Floor)
→
APx555 Reference Lab
Audio Precision APx555 Reference Noise Floor
→
Procurement & License
Request WinAudioMLS & Hardware Quote
→