SoX cheat sheet

Definition of dBFS

The definition of the dBFS unit of measurement is notably ambiguous, especially when dealing with RMS quantities.

AES17-2015 3.12.1 and IEC 61606-3:2008 3.4 both define 0 dBFS as the RMS amplitude of a full-scale sine wave.

The SoX stats filter uses a different definition, where the maximum sample value is defined as 0 dB, and the RMS value is derived from that. Thus, a full-scale sine wave has an RMS amplitude of -3 dB according to SoX:

$ sox --null --null synth 5 sine 997 stats
Pk lev dB      -0.00
RMS lev dB     -3.01

Thus, to get to the dBFS value, one must add 3.01 dB to the RMS level that SoX reports.

Generate REW-compatible WAV files

Use the --type wavpcm output option.

Generate a Dirac Pulse (unit impulse response)

Especially useful to study the effects of a digital filter.

sox --null dirac.wav trim 0 1s dcshift 0.9 pad 1 1

ITU-R BS.468-4 weighting filter

The following sox filter chain will closely match the filter specified in BS.468:

sox ... --rate 48000 ... \
	$(./generate_fir_filter.py \
		--frequency-response-spec-file=./bs468.txt \
		--dc-gain-db=-Inf --nyquist-gain-db=-Inf --antisymmetric \
		--sample-rate-hz=48000 --taps=63 --print-sox-fir)

When using such a filter for AES17-2015/IEC 61606-3:2008 measurements, don't forget to add 5.6 dB of attenuation as specified in these standards:

gain -5.6

Dynamic range measurement

The following procedure is intended to measure dynamic range as defined in AES17-2015 6.4.1 and IEC 61606-3:2008 6.2.3.3.

Assumptions

• The EUT is reasonably linear up to 0 dBFS.
  • If it's not, AES17 and IEC 61606-3 diverge, because the AES17 6.2.1 maximum input level and maximum output level, used as reference levels, will be less than 0 dBFS. In contrast, IEC 61606-3 always uses a reference level of 0 dBFS.
• The difference between the test signal frequency and the measured signal frequency is negligible.
  • If it's not, AES17 and IEC 61606-3 diverge, because AES17 specifies a notch filter set to the test signal frequency. In contrast, IEC 61606-3 5.6.3.1 specifies an auto-tuning filter.
• The upper band-edge frequency is 20 kHz.

Test signal

sox --null --bits 16 --rate 48000 997-60-16-48000.wav synth 10 sine 997 gain -60

Measurement filter

The following filters should be applied on the measured signal:

• AES17 5.2.5 low-pass filter
• AES17 5.2.8 or IEC 61606-3 5.6.3.2.6 notch/band-reject filter
• AES17 5.2.7 or IEC 61606-3 5.6.3.2.9 weighting filter

The following command implements all of the above:

sox ... --null \
	rate -v 48000 \
	highpass 4 highpass 4 \
	bandreject 997 2q \
	$(./generate_fir_filter.py \
		--frequency-response-spec-file=./bs468.txt \
		--dc-gain-db=-Inf --nyquist-gain-db=-Inf --antisymmetric \
		--sample-rate-hz=48000 --taps=63 --print-sox-fir) gain -5.6 \
	trim 1 -1 stats

Note: the purpose of the highpass filter is to remove any DC offset, which can wreak havoc in this measurement.

Note: trim 1 -1 removes invalid data at the beginning and end of the signal that is caused by filter discontinuities. This invalid data can throw off the final calculation.

Interpreting the results

The RMS lev dB value from the above command shall be compared with the value from a 997 Hz full-scale sine wave measurement. The difference between the two is the dynamic range in dB CCIR-RMS.

For convenience, this value can be interpreted as a number of bits simply by dividing by 6. (This is because adding one bit doubles the dynamic range, and a factor of two is 6 dB.) For example, the dynamic range of TPDF-dithered 16/24-bit digital audio is 96/144 dB CCIR-RMS, and the above procedure will confirm this. If sophisticated noise shaping is used instead of TPDF dither, this measurement will reflect the improved perceived dynamic range thanks to the weighting filter. For example, the measured value with SoX dither -s -p 16 is 101 dB CCIR-RMS.

THD+N vs output level measurement

The following procedure is intended to measure THD+N vs output level as defined in AES17-2015 6.3.4 and IEC 61606-3:2008 6.2.2.3.

Assumptions

• The difference between the test signal frequency and the measured signal frequency is negligible.
  • If it's not, AES17 and IEC 61606-3 diverge, because AES17 specifies a notch filter set to the test signal frequency. In contrast, IEC 61606-3 5.6.3.1 specifies an auto-tuning filter.
• The upper band-edge frequency is 20 kHz.

Test signal

sox --null --bits 16 --rate 48000 thdn-test.wav \
    synth 40 sine 997 \
    synth exp amod 0.025 0 0 0 45

Recorded signal analysis

./correlate.py \
    --reference-wav-file=thdn-test.wav \
    --test-wav-file=thdn-recorded.wav \
    --aligned-wav-file=thdn-aligned.wav
sox thdn-aligned.wav --bits 32 thdn-filtered.wav \
    highpass 4 highpass 4 \
    bandreject 997 4.3q bandreject 997 4.3q bandreject 997 4.3q \
    remix 1 trim 1 -1
sox thdn-aligned.wav thdn-reference.wav trim 1 -1
./plot_amplitude.py \
    --reference-wav-file=thdn-reference.wav \
    --wav-file=thdn-filtered.wav \
    --relative --window-size-seconds=0.4 --against-normalized-amplitude \
    --x-label 'Normalized signal level (dB)' --y-label 'THD+N ratio (dB)'

The above plot_amplitude command shows THD+N, relative to the total level of the measured signal. To plot absolute N+D, remove --relative.

Note: the purpose of the highpass filter is to remove any DC offset, which can wreak havoc in this measurement.

Note: trim 1 -1 removes invalid data at the beginning and end of the signal that is caused by filter discontinuities. This invalid data can cause spurious outliers at the ends of the plot.

Example results

Normal result where thdn-recorded.wav is a copy of thdn-test.wav:

sox thdn-test.wav thdn-recorded.wav

Performance in this example is limited by the 16-bit dithering noise floor of the test signal.

Linearity measurement

This procedure measures the amplitude linearity of the EUT; i.e. its low-level signal accuracy. It is especially effective at catching broken digital audio chains, especially dithering issues and poorly designed DACs.

This measurement can be seen as complementary to dynamic range measurements: while dynamic range measurements throw out the test signal and focuses on the noise, this linearity measurement does the exact opposite and throws out the noise to focus on the test signal.

Test signal

sox --null --bits 16 --rate 48000 linearity-test.wav \
    synth 100 sine 997 \
    synth exp amod 0.01 0 0 0 60

Recorded signal analysis

./correlate.py \
    --reference-wav-file=linearity-test.wav \
    --test-wav-file=linearity-recorded.wav \
    --aligned-wav-file=linearity-aligned.wav
sox linearity-test.wav --bits 32 linearity-reference.wav bandpass 997 200q bandpass 997 200q remix 1
sox linearity-aligned.wav --bits 32 linearity-filtered.wav bandpass 997 200q bandpass 997 200q remix 1
./plot_amplitude.py \
    --reference-wav-file=linearity-reference.wav \
    --wav-file=linearity-filtered.wav \
    --against-normalized-amplitude --relative --center --window-size-seconds=1

Example results

Ideal result where linearity-recorded.wav is a copy of linearity-test.wav:

sox linearity-test.wav linearity-recorded.wav

Normal result where linearity-recorded.wav is a redithered copy of linearity-test.wav.

sox linearity-test.wav linearity-recorded.wav gain -1

Notice how the analyser can still resolve down to around -110 dBFS despite the presence of additional 16-bit dithering noise.

Abnormal result where linearity-recorded.wav is a copy of linearity-test.wav that went through a gain filter without being redithered afterwards:

sox linearity-test.wav --no-dither linearity-recorded.wav gain -10

Adjusting the test parameters

• The test signal length determines time resolution. When changing the test signal duration, make sure to change the exponential amplitude modulation length by the inverse factor. For example, for a quicker 20-second test, change 100 to 20 and 0.01 to 0.05.
• Use the plot_amplitude --window-size-seconds parameter to adjust the smoothing of the resulting plot. Higher values make for a more readable plot but hides uncertainty in the data points.
• The amplitude lower bound of the test signal is -120 dBFS in the command above. It is twice the parameter 60. For example, to test down to -160 dBFS, change 60 to 80.
• The purpose of the analysis band-pass filter is to fish the test signal out of the dithering noise, as well as any additional noise produced by the EUT. The filter parameters determine the lower amplitude bound at which the analyser itself "bottoms out" (i.e. becomes unable to distinguish the test signal from the surrounding noise). That limit appears as a vertical "hockey stick" on the left of the resulting plot. The above command is able to resolve down to around 15 dB below a TPDF dither noise floor with around 0.5 dB accuracy; so, for example, it can measure down to around -110 dBFS (~18 bits) in the presence of 16-bit dithering noise. The absolute limit of the analyser is around -160 dBFS, presumably due to computation accuracy.