Mosqito.NET

A complete, native C# (.NET 8) port of the MoSQITo Python psychoacoustics library.

Mosqito.NET implements the same sound quality metrics as the original Python library — loudness, roughness, sharpness, tonality, and speech intelligibility — with a flat Sq.* façade that mirrors MoSQITo's public API.

Acknowledgements

Heartfelt thanks to the authors, maintainers, and contributors of the original MoSQITo Python library — in particular the team at Eomys — for their open, well-documented reference implementations of ISO 532-1, ECMA-418-2, DIN 45692, ANSI S3.5, ECMA-74, and the Daniel & Weber roughness model. Mosqito.NET would not exist without their work, and all scientific credit for the underlying algorithms belongs to them.

Disclaimer — AI-assisted port

Mosqito.NET was produced largely with AI assistance (Claude), working from the original Python MoSQITo sources as the reference implementation. The test suite (65 ported unit tests plus 168 head-to-head comparisons against Python golden outputs, all passing) is the primary guarantor of correctness, but users should still validate results against their own reference data before relying on Mosqito.NET in production or safety-critical contexts. Bug reports and corrections are very welcome.

Attribution, trademark, and endorsement

• Mosqito.NET is an independent port and is not affiliated with or endorsed by the MoSQITo project or Eomys.
• "MoSQITo" is a name used by the original project; "Mosqito.NET" refers only to this C# port and is not a claim over the original name.
• The Apache-2.0 license (§6) does not grant trademark rights; this port is licensed under Apache-2.0 for the code only.

Standards implemented

Standard	Metric
ISO 532-1:2017	Zwicker loudness (stationary + time-varying)
ECMA-418-2:2022	Loudness (hearing model), Roughness
DIN 45692:2009	Sharpness (DIN, Bismarck, Aures, Fastl weightings)
Daniel & Weber (1997)	Roughness (time-domain and frequency-domain)
ECMA-74 Annex D	Tone-to-Noise Ratio (TNR), Prominence Ratio (PR)
ANSI S3.5-1997	Speech Intelligibility Index (SII)
IEC 61260	N-th octave spectral analysis

Quick start

using Mosqito;
using Mosqito.Io;

// Load a WAV file
var (signal, fs) = WavLoader.Read("my_signal.wav");

// Zwicker stationary loudness (ISO 532-1)
var loudness = Sq.LoudnessZwst(signal, fs);
Console.WriteLine($"N = {loudness.N:F3} sone");

// DIN sharpness
double S = Sq.SharpnessDinSt(signal, fs);
Console.WriteLine($"S = {S:F3} acum");

// ECMA-418-2 roughness
var (R, RTime, RSpec, bark, time) = Sq.RoughnessEcma(signal, fs);
Console.WriteLine($"R = {R:F3} asper");

// Tone-to-noise ratio (ECMA-74)
var tnr = Sq.TnrEcmaSt(signal, fs);
Console.WriteLine($"T_total = {tnr.TTotal:F3}");

API overview

All functions are available through the Sq static façade class. Each function also has a strongly-typed class in the corresponding Mosqito.SqMetrics.* namespace.

Sound level meter

Method	Description
Sq.CompSpectrum	Complex (amplitude or dB) FFT spectrum
Sq.NoctSpectrum	N-th octave spectrum in dB
Sq.NoctSynthesis	N-th octave synthesis (amplitude bands)
Sq.FreqBandSynthesis	Arbitrary frequency-band integration

Loudness

Method	Description
Sq.LoudnessZwst	ISO 532-1 stationary loudness from time-domain signal
Sq.LoudnessZwstFreq	ISO 532-1 stationary loudness from spectrum
Sq.LoudnessZwstPerSeg	ISO 532-1 stationary loudness, segmented
Sq.LoudnessZwtv	ISO 532-1 time-varying loudness
Sq.LoudnessEcma	ECMA-418-2 loudness
Sq.EqualLoudnessContours	ISO 226 equal-loudness contours
Sq.SoneToPhon	Sone ↔ phon conversion

Roughness

Method	Description
Sq.RoughnessDw	Daniel & Weber roughness from time-domain signal
Sq.RoughnessDwFreq	Daniel & Weber roughness from spectrum
Sq.RoughnessEcma	ECMA-418-2 roughness

Sharpness

Method	Description
Sq.SharpnessDinSt	DIN 45692 stationary sharpness
Sq.SharpnessDinTv	DIN 45692 time-varying sharpness
Sq.SharpnessDinPerSeg	DIN 45692 sharpness, segmented
Sq.SharpnessDinFreq	DIN 45692 sharpness from spectrum

Tonality

Method	Description
Sq.TnrEcmaSt	ECMA-74 Tone-to-Noise Ratio, stationary
Sq.TnrEcmaFreq	ECMA-74 TNR from spectrum
Sq.TnrEcmaPerSeg	ECMA-74 TNR, segmented
Sq.PrEcmaSt	ECMA-74 Prominence Ratio, stationary
Sq.PrEcmaFreq	ECMA-74 PR from spectrum
Sq.PrEcmaPerSeg	ECMA-74 PR, segmented

Speech intelligibility

Method	Description
Sq.SiiAnsi	ANSI S3.5 Speech Intelligibility Index from signal
Sq.SiiAnsiFreq	ANSI S3.5 SII from spectrum

Performance

All public APIs were accelerated via Parallel.For, SIMD (System.Numerics.Vector<double>), ArrayPool<T>, plan caches, and allocation removal across two optimisation passes. Numbers below are from BenchmarkDotNet on an AMD Ryzen 7 4800H (8 cores / 16 threads), .NET 8 AVX2, comparing the unoptimised initial implementation to the current build.

Loudness

Benchmark	Before	After	Speedup	Alloc Δ
LoudnessZwst (synthetic)	29.94 ms	5.70 ms	5.3×	−65%
LoudnessZwst (WAV file)	312.66 ms	61.45 ms	5.1×	−69%
LoudnessZwstFreq (WAV file)	572.06 ms	3.57 ms	160×	−99%
LoudnessZwtv (synthetic)	39.65 ms	15.08 ms	2.6×	−7%
LoudnessZwtv (WAV file)	475.30 ms	201.90 ms	2.4×	—
LoudnessEcma (synthetic)	84.96 ms	34.29 ms	2.5×	−44%
LoudnessEcma (WAV file)	1,047.49 ms	358.89 ms	2.9×	−18%

LoudnessZwstFreq gains are dominated by plan-caching the critical-band filter-bank design (previously recomputed from scratch on every call).

Roughness

Benchmark	Before	After	Speedup	Alloc Δ
RoughnessDw (synthetic)	3,732.9 ms	687.4 ms	5.4×	—
RoughnessDw (WAV file)	11,338.7 ms	2,216.8 ms	5.1×	—
RoughnessDwFreq (synthetic)	749.3 ms	150.7 ms	5.0×	—
RoughnessDwFreq (WAV file)	772.8 ms	150.8 ms	5.1×	—
RoughnessEcma (synthetic)	1,977.0 ms	464.4 ms	4.3×	−24%
RoughnessEcma (WAV file)	4,676.5 ms	979.3 ms	4.8×	−28%

Sharpness

Benchmark	Before	After	Speedup	Alloc Δ
SharpnessDinSt (synthetic)	30.73 ms	6.14 ms	5.0×	−63%
SharpnessDinSt (WAV file)	154.82 ms	30.31 ms	5.1×	−61%
SharpnessDinTv (synthetic)	40.95 ms	14.79 ms	2.8×	−6%
SharpnessDinTv (WAV file)	217.34 ms	84.07 ms	2.6×	−9%

Tonality (TNR / PR) and Speech Intelligibility (SII)

Single-call benchmarks are FFT-dominated; the parallel gains apply to the PerSeg variants on longer signals. Allocation reductions from stackalloc locals and removed LINQ are present in all call paths.

Benchmark	Before	After	Alloc Δ
TnrEcmaSt (synthetic)	34.17 ms	34.20 ms	−18%
TnrEcmaSt (WAV file)	33.92 ms	35.45 ms	−19%
PrEcmaSt (synthetic)	34.15 ms	34.44 ms	−18%
PrEcmaSt (WAV file)	34.54 ms	34.36 ms	−19%
SiiAnsi (synthetic)	31.30 ms	30.19 ms	−20%
SiiAnsi (WAV file)	240.46 ms	239.31 ms	−15%

Dependencies

• MathNet.Numerics 5.0 — numerical routines (FFT, interpolation, filtering)
• NAudio 2.2 — WAV file I/O

Test results

The library ships a comprehensive test suite with 233 tests, all passing (as of 2026-04-16, run time ~7 minutes).

Test Run Successful.
Total tests:  233
Passed:       233
Failed:         0

Test tiers

Tier 1 — Ported MoSQITo unit tests (65 cases)

Direct ports of the original Python MoSQITo test suite, validating each module against ISO/ECMA/DIN reference values.

Test class	Cases	What it validates
LoudnessZwstTests	5	ISO 532-1 stationary loudness — time-domain, freq-domain, per-segment, 44.1 kHz resample, 1/3-octave input
LoudnessZwtvTests	3	ISO 532-1 time-varying loudness — dimensions, peak detection, time-average convergence
LoudnessEcmaTests	3	ECMA-418-2 loudness — dimensions, equal-loudness contour, 1 kHz vs 5 kHz comparison
SharpnessDinTests	6	DIN 45692 sharpness — all 4 weightings, freq-domain path, per-segment
RoughnessDwTests	3	Daniel-Weber roughness — AM tone ~1 asper (time & freq domain), dimensions
RoughnessEcmaTests	3	ECMA-418-2 roughness — AM tone 1 kHz/70 Hz/60 dB ~1 asper, R_time positive, dimensions
TnrEcmaTests	5	ECMA TNR — 2-tone detection, t_total value, per-segment, freq path, non-prominent tones
PrEcmaTests	5	ECMA PR — same 5 patterns as TNR
SiiAnsiTests	4	ANSI S3.5 SII — octave-method reference 0.504, level vs freq entry equivalence, all methods
NoctTests	8	N-th octave spectrum & synthesis — 1/3 & full octave, 94 dB SPL @ 1 kHz, 3 dB band addition
TimeSegmentationTests	6	I/O segmentation — block count, shape, overlap, monotonic time axis, ECMA 24-block mode
SqFacadeTests	14	Facade smoke tests for every public Sq.* method

Tier 2 — Head-to-head Python golden tests (168 cases)

Each of the 24 public functions is run against 7 WAV files and compared to outputs pre-generated by the original Python MoSQITo library. Tolerance is rtol = 1% for psychoacoustic metrics, atol = 0.5 dB for dB-domain arrays.

Functions covered: loudness_zwst, loudness_zwst_freq, loudness_zwst_perseg, loudness_zwtv, loudness_ecma, sharpness_din_st, sharpness_din_tv, sharpness_din_perseg, sharpness_din_freq, roughness_dw, roughness_dw_freq, roughness_ecma, tnr_ecma_st, pr_ecma_st, tnr_ecma_freq, pr_ecma_freq, tnr_ecma_perseg, pr_ecma_perseg, sii_ansi, sii_ansi_freq, noct_spectrum, noct_synthesis, comp_spectrum, freq_band_synthesis

WAV signals used:

Signal	Sample rate	Calibration
white_noise_200_2000_Hz_stationary	48 kHz	0.01
white_noise_442_1768_Hz_stationary	48 kHz	0.01
white_noise_442_1768_Hz_varying	48 kHz	0.01
broadband_570	48 kHz	1.0
Test signal 3 (1 kHz 60 dB)	44.1 kHz	2√2
Test signal 5 (pinknoise 60 dB)	48 kHz	2√2
Test signal 10 (tone pulse 1 kHz 10 ms 70 dB)	48 kHz	2√2

Running the tests

dotnet test

Note: the Daniel-Weber roughness head-to-head tests are computationally intensive (~2–3 minutes for the long WAV files). The full suite completes in approximately 7 minutes.

License

Mosqito.NET is derived from MoSQITo (Apache License 2.0) and is distributed under the same license. See NOTICE.md for full attribution.