Seabass - Python Implementation of the 2f Model

Seabass is a Python implementation of the 2f-model developed by the Internaltional Audio Laboratories Erlangen. It is a model for estimating subjective quality of separated audio source signals, specifically the mean MUSHRA score. The part of this code, particularly the computation of the Model Output Variables are based on the MATLAB implementation, PQEvalAudio, by Peter Kabal. See here for more details about the 2f-model.

Installation

git clone this-repository
cd 2f-model
pip install -e .

Usage

The computation of MOVs requires that the audio signal are scaled to have amplitudes in the range [-32768, 32767]. By default, seabass handles this automatically when using the CLI mode or estimate_mean_mushra_score_file, but not always when using estimate_mean_mushra_score.

When reading from files, soundfile reads the header and normalize it to floating point in the range [-1, 1). seabass then scales them up by 32768.

CLI

seabass --test-path /path/to/test-file.wav \
    --ref-path /path/to/reference-file.wav \
    --output-path /path/to/output.json

Python

Evaluating directly from files

from seabass.core.models.mms import estimate_mean_mushra_score_file

mms = estimate_mean_mushra_score_file(test_path, ref_path)

print(f"Estimated MMS: {mms:3.2f}")

Evaluating from loaded signals

When using loaded audio, make sure that your audio signals are either

• in the range [-32768, 32767] and pass the argument signal_dtype='int16' to the function, or
• in the range [-1, 1) and pass the argument signal_dtype='double' to the function.

If your audio signals are in the range [-1, 1), they will be multiplied by 32768 to scale them to [-32768, 32768). If this is not what you want, you can scale them yourself before passing them to the function with signal_dtype='int16'.

import soundfile as sf
from seabass.core.models.mms import estimate_mean_mushra_score

ref_signal, fs_r = sf.read(ref_path, always_2d=True)
test_signal, fs_t = sf.read(test_path, always_2d=True)

# soundfile returns (n_samples, n_channels) but seabass expectes (n_channels, n_samples)
ref_signal = ref_signal.T 
test_signal = test_signal.T

assert fs_r == fs_t == 48000
# PEAQ only works with 48 kHz signals
# if your signals are not 48 kHz, resample them

mms = estimate_mean_mushra_score(test_signal, ref_signal)

print(f"Estimated MMS: {mms:3.2f}")