SPB Git

Repository Public License

dsp-validation.md 8.2 KB

# DSP Validation Report

Date: 2026-11-20 · Suite: ORIXAudioCoreTests · Result: 83 tests in 14 suites, all passed · Platform: macOS 27, Apple M5 Max · Run with: Scripts/test.sh core

# Method

Each measurement is compared with a synthetic signal whose correct answer is known from theory. Signals are generated deterministically (SignalGenerator, seeded noise), at 48 kHz unless stated. A tolerance is fixed in the test; the test passes only inside it. The tolerances below are the ones asserted, not observed errors — the suite reports pass or fail, not the residual.

This validates the mathematics. It says nothing about microphones, rooms or devices.

# Spectrum

Test Signal Expected Tolerance
Window gains six windows, 4096 points mean and mean-square of the coefficients equal the declared coherent and noise power gains 1e-6
Equivalent noise bandwidth — rectangular 1.0, Hann 1.5, Hamming 1.3628, Blackman–Harris 2.0044, flat top 3.7702 1e-9 to 1e-2
Amplitude scaling bin-centred sine, A = 0.5, N = 256, 1024, 4096, 16 384, all six windows peak at the right bin, reading 0.5 0.001
Power scaling bin-centred sine, A = 0.8 0.32 (A²/2) 0.001
DC and Nyquist constant 0.25; alternating ±0.3 0.25; 0.3 (not doubled) 1e-4
Flat-top accuracy sine 0, ¼ and ½ bin off centre 0 dBFS 0.02 dB
Hann scalloping sine ½ bin off centre 1.42 dB loss 1.3–1.5 dB
Leakage off-bin sine, 40 bins away rectangular above −45 dB, Hann below −95 dB —
Resolution two tones 4 bins apart both read ≥ 0.39 of 0.4, the bin between < 0.01 —
PSD white noise, σ = 0.1, 200 frames integral equals σ²; level equals σ²/(fs/2) 3 %; 5 %
Numerical safety silence; constant full scale finite values, floor ≤ −150 dB —

# Spectrum analyser

Test Signal Expected Tolerance
Dominant frequency sines at 61.7, 440, 1234.5, 9876.5 Hz, N = 8192 the frequency; level −10.46 dBFS 0.5 Hz; 0.3 dB
Silence zeros no dominant frequency —
Band power sines at 40, 120, 350, 1000, 3000, 5000, 10 000 Hz, A = 0.5 0.125 in the expected band, < 0.002 elsewhere 3 %
Peak hold full-scale tone then silence, 10 dB/s −5 dB after 0.5 s 0.05 dB
Smoothing tone then silence, factor 0.5 −3.01 dB 0.05 dB
Peaks 500, 2000, 7000 Hz at 0.5, 0.25, 0.1 three peaks at those frequencies and levels 2 Hz; 0.3 dB
Log axis single-bin peak preserved after resampling to 200 columns exact

# Spectrogram

Test Signal Expected Tolerance
Timing 1875 Hz burst from 1.0 to 1.5 s onset and end at those times; level −6.02 dBFS inside 2 hops; 0.3 dB
Chirp 500 → 8000 Hz over 2 s ridge follows the instantaneous frequency 2 bins
Impulse one sample energy only in columns whose window contains it exact
Streaming pink noise pushed in uneven chunks identical to one-shot exact
Raster 6 kHz tone, 64 rows, 0–24 kHz linear brightest row at ¼ height ±1 row
Palettes five palettes luminance non-decreasing; range > 150 —

# Time domain

Test Expected Tolerance
Sine, A = 0.5, 1 kHz peak 0.5; RMS 0.3536; crest 3.01 dB; 2000 crossings/s 1e-4; 0.01 dB; 2 /s
White noise, σ = 0.1 RMS −20 dBFS; crossings fs/2; crest 11–15 dB 0.1 dB; 2 %
Clipped 100 Hz sine 200 clipping events; a clean full-scale sine gives none exact
Silence from 2 to 4 s one silence at those times 0.4 s
DC 0.2 offset 0.2 1e-5
Chunking identical metrics however the signal is split exact
One hour of a −20 dBFS sine RMS −23.0103 dBFS 0.001 dB

# Spectral descriptors

Signal Expected Tolerance
Pure tones at 250, 1000, 6000 Hz centroid and rolloff at the tone; bandwidth < 40 Hz; flatness < 0.001 3 Hz; one bin
White noise centroid fs/4; bandwidth 24 000/√12 Hz; rolloff 0.85 × 24 kHz; flatness e^(−γ) ≈ 0.561 300 Hz; 500 Hz; 0.03
Pink noise equal power in the 250–500 Hz and 2–4 kHz octaves 1 dB
Two tones, 0.4 and 0.2 band levels and an 80 / 20 power split 0.2 dB; 1 %

# Loudness

Test Signal Expected Tolerance
K-weighting fs = 48 kHz the coefficients tabulated in BS.1770 1e-8
Reference tone 0 dBFS, 997 Hz, mono, at 44.1, 48 and 96 kHz −3.01 LKFS 0.05
EBU Tech 3341 case 1, 2 stereo 1 kHz at −23 and −33 dBFS, 20 s −23.0 and −33.0 LUFS 0.1 LU
Case 3 (relative gate) −36, −23, −36 dBFS for 10, 60, 10 s −23.0 LUFS 0.1 LU
Case 4 (both gates) −72, −36, −23, −36, −72 dBFS −23.0 LUFS 0.1 LU
Case 5 −26, −20, −26 dBFS for 20, 20.1, 20 s −23.0 LUFS 0.1 LU
EBU Tech 3342 cases 1, 2 −20 then −30 dBFS; −20 then −15 dBFS 10 LU; 5 LU 1 LU
True peak fs/4 sine at 45° (samples at ±0.707) 0 dBTP 0.3 dB
True peak 440 Hz at −6 dBFS −6 dBTP 0.1 dB
No reading silence; 0.2 s of signal integrated loudness is absent —

Scope of the loudness claim. The signals were synthesised to the published descriptions of the EBU test cases; the official EBU audio files were not used, and only the cases listed were run. Mono and stereo are implemented; surround channel weights are not. The correct statement is implements ITU-R BS.1770-4 and EBU Tech 3342 loudness range, verified against synthesised reference signals. It is not a certification of EBU R 128 compliance.

# Waveform, voice activity, capture

Test Expected
Bucket statistics min, max and RMS of a sine within 0.005
Colour band tones at 45, 150, 350, 1000, 3000, 5000 and 9000 Hz land in bands 0–6 with more than 90 % of band power; silence has no band
Band follows content 100 Hz then 8 kHz: dominant band changes from 1 to 6
Pyramid and cache extremes preserved to the top level; serialisation round-trips; junk data rejected
Voice activity three 1 s bursts over a −60 dBFS floor are found within 80 ms of their starts; steady noise alone yields nothing
CAF master 16-bit, 24-bit and float files read back by Core Audio within one quantisation step
Interrupted master unclosed, with a torn byte: 96 000 frames reported and read back
Ring buffer two million samples across threads in order; overflow counted

# Editing and effects

Test Expected Tolerance
Render equals the edit list delete, insert silence and −6.02 dB on a range reproduce the source sample for sample 2 LSB at 24 bits
Fades, normalise, channel modes raised-cosine midpoints; peak lands on −1 dBFS; left, right and mono values 1e-3; 0.01 dB
High-pass 200 Hz −3.01 dB at 200 Hz; −24.1 dB at 100 Hz; 0 dB at 2 kHz 0.15; 0.6; 0.05 dB
Low-pass 4 kHz −3.01 dB at 4 kHz; below −22 dB at 8 kHz 0.15 dB
Equaliser +6 dB at 30 Hz, −4 dB at 1.5 kHz, +5 dB at 16 kHz 0.3; 0.1; 0.3 dB
Compressor static curve values; a steady tone 12 dB over a 3:1 threshold drops 8 dB 1e-9; 0.3 dB
Denoiser, empty profile output equals input after latency compensation 2e-4
Noise reduction at 70 % noise floor falls more than 10 dB; a 1 kHz tone keeps its frequency and level; onset stays at 3.0 s 1 Hz; 0.5 dB; 5 ms
Lossless encoders WAV 16/24, AIFF 16, CAF 24, Apple Lossless round-trip one quantisation step
AAC every combination of 8–96 kHz, mono/stereo and 32–256 kbps requested encodes and decodes duration within 0.15 s
Resampling 48 → 24 kHz tone frequency and duration preserved 2 Hz; 0.01 s

# What is not validated

  • Anything acoustic: microphone frequency response, noise, automatic gain.
  • Absolute level. All figures are relative to digital full scale.
  • Speech-activity accuracy on real speech; the detector was checked on noise bursts only.
  • The usefulness of the findings on real recordings. One rule was reworded during development after it fired on dense music.
  • Noise reduction on real recordings; it was verified on white noise with a tone.
Files main
Menu

Theme