Source code for bob.bio.spear.extractor.CQCCFeatures

#!/usr/bin/env python
# vim: set fileencoding=utf-8 :
# Pavel Korshunov <pavel.korshunov@idiap.ch>
# Wed 23 Nov 17:13:22 CEST 2016
#


from __future__ import print_function

import bob.io.base
import numpy

import logging
logger = logging.getLogger("bob.bio.spear")
logger.setLevel(logging.DEBUG)

from bob.bio.base.extractor import Extractor
from bob.bio.base.preprocessor import Preprocessor


class CQCCFeatures(Preprocessor, Extractor):
    """
    This class should be used as a preprocessor (converts matlab data into HDF5) and an extractor (reads saved data)

    Converts pre-computed with Matlab CQCC features into numpy array suitable for Bob-based experiments.
    CQCC features are obtained using CQCC Matlab toolkit from http://audio.eurecom.fr/content/software
    The features are originally proposed in the following paper:
    Todisco, Massimiliano; Delgado, Héctor; Evans, Nicholas
    "Articulation rate filtering of CQCC features for automatic speaker verification", INTERSPEECH 2016,
    Annual Conference of the International Speech Communication Association, September 8-12, 2016, San Francisco, USA
    """

    def __init__(
            self,
            split_training_data_by_client=False,
            features_mask=numpy.zeros(90),  # mask of which features to read
            **kwargs
    ):
        # call base class constructor with its set of parameters
        Preprocessor.__init__(self, read_original_data=self.read_matlab_files, **kwargs)
        Extractor.__init__(self, requires_training=False, split_training_data_by_client=split_training_data_by_client,
                           **kwargs)
        self.features_mask = features_mask

    def read_matlab_files(self, biofile, directory, extension):
        """
        Read pre-computed CQCC Matlab features here
        """
        import bob.io.matlab
        # return the numpy array read from the data_file
        data_path = biofile.make_path(directory, extension)
        return bob.io.base.load(data_path)

    def __call__(self, input_data, annotations):
        """
        When this function is called in the capacity of Preprocessor, we apply feature mask to the features.
        When it is called as an Extractor, we assume that we have correct CQCC features already,
        so we can pass them on to the classifier
        """

        features = input_data
        # features mask cannot be large the the features themselves
        assert(self.features_mask.shape[0] < input_data.shape[0])
        if self.features_mask.shape[0] < input_data.shape[0]:  # apply the mask
            features = input_data[self.features_mask]

        # transpose, because of the way original Matlab-based features are computed
        return numpy.transpose(features)


cqcc_reader = CQCCFeatures()