pesq

PESQ (Perceptual Evaluation of Speech Quality) Wrapper for Python [Updated for P.862 Corrigendum 2 (03/18)]

Description

This code is an updated version of /ludlows/pesq/ which implements the Corrigendum 2 of the ITU-T P.862 recommendation (PESQ). The correction addresses the under-prediction of subjective scores (by 0.8 MOS on average) by correcting the level of the loudness model.

This code is designed for numpy array specially.

Requirements

C compiler
numpy
cython

Install with pip

# PyPi Repository
$ pip install pesq

# The Latest Version
$ pip install https://github.com/ludlows/python-pesq/archive/master.zip

Usage for narrowband and wideband Modes

Please note that the sampling rate (frequency) should be 16000 or 8000 (Hz).

A sample rate of 8000 Hz is supported only in narrowband mode.

The code supports error-handling behaviors now.

def pesq(fs, ref, deg, mode='wb', on_error=PesqError.RAISE_EXCEPTION):
    """
    Args:
        ref: numpy 1D array, reference audio signal 
        deg: numpy 1D array, degraded audio signal
        fs:  integer, sampling rate
        mode: 'wb' (wide-band) or 'nb' (narrow-band)
        on_error: error-handling behavior, it could be PesqError.RETURN_VALUES or PesqError.RAISE_EXCEPTION by default
    Returns:
        pesq_score: float, P.862.2 Prediction (MOS-LQO)
    """

Once you select PesqError.RETURN_VALUES, the pesq function will return -1 when an error occurs.

Once you select PesqError.RAISE_EXCEPTION, the pesq function will raise an exception when an error occurs.

It now supports the following errors: InvalidSampleRateError, OutOfMemoryError,BufferTooShortError,NoUtterancesError,PesqError(other unknown errors).

from scipy.io import wavfile
from pesq import pesq

rate, ref = wavfile.read("./audio/speech.wav")
rate, deg = wavfile.read("./audio/speech_bab_0dB.wav")

print(pesq(rate, ref, deg, 'wb'))
print(pesq(rate, ref, deg, 'nb'))

Usage for multiprocessing feature

def pesq_batch(fs, ref, deg, mode='wb', n_processor=None, on_error=PesqError.RAISE_EXCEPTION):
    """
   Running `pesq` using multiple processors
    Args:
        on_error:
        ref: numpy 1D (n_sample,) or 2D array (n_file, n_sample), reference audio signal
        deg: numpy 1D (n_sample,) or 2D array (n_file, n_sample), degraded audio signal
        fs:  integer, sampling rate
        mode: 'wb' (wide-band) or 'nb' (narrow-band)
        n_processor: cpu_count() (default) or number of processors (chosen by the user) or 0 (without multiprocessing)
        on_error: PesqError.RAISE_EXCEPTION (default) or PesqError.RETURN_VALUES
    Returns:
        pesq_score: list of pesq scores, P.862.2 Prediction (MOS-LQO)
    """

This function uses multiprocessing features to boost time efficiency.

When the ref is an 1-D numpy array and deg is a 2-D numpy array, the result of pesq_batch is identical to the value of [pesq(fs, ref, deg[i,:],**kwargs) for i in range(deg.shape[0])].

When the ref is a 2-D numpy array and deg is a 2-D numpy array, the result of pesq_batch is identical to the value of [pesq(fs, ref[i,:], deg[i,:],**kwargs) for i in range(deg.shape[0])].

Correctness

The correctness is verified by running samples in the audio folder.

PESQ computed by this code in wideband mode is 1.5128041505813599 1.0832337141036987 [due to Corrigendum 2]

PESQ computed by this code in narrowband mode is 1.6072081327438354

Note

Sampling rate (fs|rate) - No default. You must select either 8000Hz or 16000Hz.

Note there is narrowband (nb) mode only when the sampling rate is 8000Hz.

The original C source code is modified.

Who is using pesq

Please click here to see these repositories, whose owners include Facebook Research, SpeechBrain, NVIDIA .etc.

Acknowledgement

The work at /ludlows/pesq was funded by the Natural Sciences and Engineering Research Council of Canada.

The work at /ludlows/pesq was also funded by the Concordia University, Montreal, Canada.