MetaSona

Psychoacoustic metrics in C, with a typed Python interface.

MetaSona packages compiled C implementations in a Python wheel to reduce computation time for repeated psychoacoustic analysis, while keeping a simple NumPy interface. The speed comes from the native implementation; the wheel makes that implementation easy to install.

The current version is experimental, particularly roughness and tonality. Standards conformance has not been established.

The metrics draw on SQAT revision e6228b789fc9. Aures tonality uses 250 ms windows and 125 ms hops. MetaSona returns every complete frame with centre timestamps; see implementation credits and differences.

Computation time

Measured September 2026 with the SQAT-aligned MetaSona wheel. Median of five warmed calls for 10 s of 48 kHz audio (1 kHz carrier, 70 Hz AM plus noise, 60 dB SPL), Ryzen 9 5950X, Windows. MetaSona: MSVC Release; MoSQITo: 1.2.1; imports and audio I/O are excluded.

These are workload timings; models and outputs differ. Native filtering and channel processing contribute to speed alongside FFTs; compiler and framing differences also affect timings.

Python

MetaSona requires Python 3.9–3.14. Install the published package from PyPI with pip:

python -m pip install metasona

Or add it to a uv project:

uv add metasona

Prebuilt wheels are published for Windows x64, Linux x64/ARM64, and macOS Intel/Apple silicon, so these installations do not require a C compiler.

To install from a source checkout instead, use python -m pip install . or uv add /path/to/metasona. Building from source requires a C11 compiler.

Pass calibrated pressure in pascals, not uncalibrated audio samples. The ECMA functions accept mono (samples,) or stereo (samples, 2) arrays; the other functions accept mono arrays. Python accepts integer sample rates from 8 to 192 kHz and resamples to 48 kHz when needed. Results include their units and read-only NumPy arrays.

For ECMA loudness and tonality together, use result = ms.ecma_tonal_analysis(pressure_pa, fs) and read result.loudness and result.tonality. Loudness/tonality require at least 304 ms of audio; roughness requires 320 ms. Representative values exclude the initial filter transient. ECMA loudness implements Section 8, whereas MoSQITo's loudness_ecma implements the Section 5 basis calculation. ECMA roughness does not apply the optional entropy weighting. Function and result-class docstrings describe output shapes, units, timing and binaural combination.

import numpy as np
import metasona as ms

# One second of a 1 kHz tone at 60 dB SPL (0.02 Pa RMS).
fs = 48_000
t = np.arange(fs) / fs
pressure_pa = np.sqrt(2) * 0.02 * np.sin(2 * np.pi * 1_000 * t)

loudness = ms.stationary_loudness(pressure_pa, fs, sound_field="free")
sharpness = ms.sharpness_din45692(loudness.specific_loudness_sone_per_bark)

print(f"Loudness: {loudness.loudness_sone:.3f} sone")
print(f"Sharpness: {sharpness.sharpness_acum:.3f} acum")

See the Python API and complete example.

C library

Requires CMake 3.21 or newer and a C11 compiler:

cmake -S . -B build -DCMAKE_BUILD_TYPE=Release
cmake --build build --config Release
cmake --install build --config Release --prefix install

Include <metasona/metasona.h> and link the installed CMake target MetaSona::metasona. Native signal functions require 48 kHz input. Use -DMS_BUILD_SHARED=OFF for a static library.

See the C API and C example.

License and credits

GPL-3.0 for the combined distribution, with retained Apache-2.0, BSD-3-Clause and MIT component notices. See third-party credits.

Acknowledgements

Developed by Jiahua Zhang with assistance from OpenAI Codex.

Developed during PhD research within the METAVISION MSCA Doctoral Network.

The European Commission is gratefully acknowledged for their support of the Horizon Europe DN METAVISION project (GA 101072415). Views and opinions expressed are however those of the authors only and do not necessarily reflect those of the European Union. The European Union cannot be held responsible for them.