BEAT - pkorshunov/vad

This algorithm is a legacy one. The API has changed since its implementation. New versions and forks will need to be updated.

This algorithm is splittable

Endpoint Groups 1

Algorithms have at least one input and one output. All algorithm endpoints are organized in groups. Groups are used by the platform to indicate which inputs and outputs are synchronized together. The first group is automatically synchronized with the channel defined by the block in which the algorithm is deployed.

Group: main

Endpoint Name	Data Format	Nature
speech	system/array_1d_floats/1	Input
labels	system/array_1d_integers/1	Output

Parameters 3

Parameters allow users to change the configuration of an algorithm when scheduling an experiment

Name	Description	Type	Default	Range/Choices
rate	Sampling rate of the speech signal	float64	16000.0	[2000.0, 256000.0]
win_length_ms	The length of the sliding processing window, typically about 20 ms	float64	20.0
win_shift_ms	The length of the overlap between neighboring windows. Typically the half of window length.	float64	10.0

###############################################################################
#                                                                             #
# Copyright (c) 2016 Idiap Research Institute, http://www.idiap.ch/           #
# Contact: beat.support@idiap.ch                                              #
#                                                                             #
# This file is part of the beat.core module of the BEAT platform.             #
#                                                                             #
# Commercial License Usage                                                    #
# Licensees holding valid commercial BEAT licenses may use this file in       #
# accordance with the terms contained in a written agreement between you      #
# and Idiap. For further information contact tto@idiap.ch                     #
#                                                                             #
# Alternatively, this file may be used under the terms of the GNU Affero      #
# Public License version 3 as published by the Free Software and appearing    #
# in the file LICENSE.AGPL included in the packaging of this file.            #
# The BEAT platform is distributed in the hope that it will be useful, but    #
# WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY  #
# or FITNESS FOR A PARTICULAR PURPOSE.                                        #
#                                                                             #
# You should have received a copy of the GNU Affero Public License along      #
# with the BEAT platform. If not, see http://www.gnu.org/licenses/.           #
#                                                                             #
###############################################################################

import numpy
import math
import scipy.signal
import bob.ap

class Mod_4Hz():
    """VAD based on the modulation of the energy around 4 Hz and the energy """
    def __init__(
            self,
            max_iterations = 10,        # 10 iterations for the
            convergence_threshold = 0.0005,
            variance_threshold = 0.0005,
            win_length_ms = 20.,        # 20 ms
            win_shift_ms = 10.,           # 10 ms
            smoothing_window = 10, # 10 frames (i.e. 100 ms)
            n_filters = 40,
            f_min = 0.0,                       # 0 Hz
            f_max = 4000,                   # 4 KHz
            pre_emphasis_coef = 1.0,
            ratio_threshold = 0.1,       # 0.1 of the maximum energy
            **kwargs
    ):
        # copy parameters
        self.max_iterations = max_iterations
        self.convergence_threshold = convergence_threshold
        self.variance_threshold = variance_threshold
        self.win_length_ms = win_length_ms
        self.win_shift_ms = win_shift_ms
        self.smoothing_window = smoothing_window
        self.n_filters = n_filters
        self.f_min = f_min
        self.f_max = f_max
        self.pre_emphasis_coef = pre_emphasis_coef
        self.ratio_threshold = ratio_threshold

def _voice_activity_detection(self, energy, mod_4hz):

n_samples = len(energy)
        threshold = numpy.max(energy) - numpy.log((1./self.ratio_threshold) * (1./self.ratio_threshold))
        labels = numpy.array(numpy.zeros(n_samples), dtype=numpy.int16)

for i in range(n_samples):
            if ( energy[i] > threshold and mod_4hz[i] > 0.9 ):
                labels[i]=1

# If speech part less then 10 seconds and less than the half of the segment duration, try to find speech with more risk
        if  numpy.sum(labels) < 2000 and float(numpy.sum(labels)) / float(len(labels)) < 0.5:
            # TRY WITH MORE RISK 1...
            for i in range(n_samples):
                if ( energy[i] > threshold and mod_4hz[i] > 0.5 ):
                    labels[i]=1

if  numpy.sum(labels) < 2000 and float(numpy.sum(labels)) / float(len(labels)) < 0.5:
            # TRY WITH MORE RISK 2...
            for i in range(n_samples):
                if ( energy[i] > threshold and mod_4hz[i] > 0.2 ):
                    labels[i]=1

if  numpy.sum(labels) < 2000 and float(numpy.sum(labels)) / float(len(labels)) < 0.5: # This is special for short segments (less than 2s)...
            # TRY WITH MORE RISK 3...
            if (len(energy) < 200 ) or (numpy.sum(labels) == 0) or (numpy.mean(labels)<0.025):
                for i in range(n_samples):
                    if ( energy[i] > threshold ):
                        labels[i]=1
        return labels

def averaging(self, list_1s_shift):
        len_list=len(list_1s_shift)
        sample_level_value = numpy.array(numpy.zeros(len_list, dtype=numpy.float))
        sample_level_value[0]=numpy.array(list_1s_shift[0])
        for j in range(2, numpy.min([len_list, 100])):
            sample_level_value[j-1]=((j-1.0)/j)*sample_level_value[j-2] +(1.0/j)*numpy.array(list_1s_shift[j-1])
        for j in range(numpy.min([len_list, 100]), len_list-100 +1):
            sample_level_value[j-1]=numpy.array(numpy.mean(list_1s_shift[j-100:j]))
        sample_level_value[len_list-1] = list_1s_shift[len_list -1]
        for j in range(2, numpy.min([len_list, 100]) + 1):
            sample_level_value[len_list-j]=((j-1.0)/j)*sample_level_value[len_list+1-j] +(1.0/j)*numpy.array(list_1s_shift[len_list-j])
        return sample_level_value

def bandpass_firwin(self, ntaps, lowcut, highcut, fs, window='hamming'):
        nyq = 0.5 * fs
        taps = scipy.signal.firwin(ntaps, [lowcut, highcut], nyq=nyq, pass_zero=False,
                                   window=window, scale=True)
        return taps

def pass_band_filtering(self, energy_bands, fs):
        energy_bands = energy_bands.T
        order = 8
        Wo = 4.
        num_taps = self.bandpass_firwin(order+1, (Wo - 0.5), (Wo + 0.5), fs)
        res = scipy.signal.lfilter(num_taps, 1.0, energy_bands)
        return res

def modulation_4hz(self, filtering_res, rate_wavsample):
        fs = rate_wavsample[0]
        win_length = int (fs * self.win_length_ms / 1000)
        win_shift = int (fs * self.win_shift_ms / 1000)
        Energy = filtering_res.sum(axis=0)
        mean_Energy = numpy.mean(Energy)
        Energy = Energy/mean_Energy

#        win_size = int (2.0 ** math.ceil(math.log(win_length) / math.log(2)))
        n_frames = 1 + (rate_wavsample[1].shape[0] - win_length) // win_shift
        range_modulation = int(fs/win_length) # This corresponds to 1 sec
        res = numpy.zeros(n_frames)
        if n_frames < range_modulation:
            return res
        for w in range(0,n_frames-range_modulation):
            E_range=Energy[w:w+range_modulation] # computes the modulation every 10 ms
            if (E_range<=0.).any():
                res[w] = 0
            else:
                res[w] = numpy.var(numpy.log(E_range))
        res[n_frames-range_modulation:n_frames] = res[n_frames-range_modulation-1]
        return res

def smoothing(self, labels, smoothing_window):
        """ Applies a smoothing on VAD"""

if numpy.sum(labels)< smoothing_window:
            return labels
        segments = []
        for k in range(1,len(labels)-1):
            if labels[k]==0 and labels[k-1]==1 and labels[k+1]==1 :
                labels[k]=1
        for k in range(1,len(labels)-1):
            if labels[k]==1 and labels[k-1]==0 and labels[k+1]==0 :
                labels[k]=0

seg = numpy.array([0,0,labels[0]])
        for k in range(1,len(labels)):
            if labels[k] != labels[k-1]:
                seg[1]=k-1
                segments.append(seg)
                seg = numpy.array([k,k,labels[k]])

seg[1]=len(labels)-1
        segments.append(seg)

if len(segments) < 2:
            return labels

curr = segments[0]
        next = segments[1]

# Look at the first segment. If it's short enough, just change its labels
        if (curr[1]-curr[0]+1) < smoothing_window and (next[1]-next[0]+1) > smoothing_window:
            if curr[2]==1:
                labels[curr[0] : (curr[1]+1)] = numpy.zeros(curr[1] - curr[0] + 1)
                curr[2]=0
            else: #curr[2]==0
                labels[curr[0] : (curr[1]+1)] = numpy.ones(curr[1] - curr[0] + 1)
                curr[2]=1

for k in range(1,len(segments)-1):
            prev = segments[k-1]
            curr = segments[k]
            next = segments[k+1]

if (curr[1]-curr[0]+1) < smoothing_window and (prev[1]-prev[0]+1) > smoothing_window and (next[1]-next[0]+1) > smoothing_window:
                if curr[2]==1:
                    labels[curr[0] : (curr[1]+1)] = numpy.zeros(curr[1] - curr[0] + 1)
                    curr[2]=0
                else: #curr[2]==0
                    labels[curr[0] : (curr[1]+1)] = numpy.ones(curr[1] - curr[0] + 1)
                    curr[2]=1

prev = segments[-2]
        curr = segments[-1]

if (curr[1]-curr[0]+1) < smoothing_window and (prev[1]-prev[0]+1) > smoothing_window:
            if curr[2]==1:
                labels[curr[0] : (curr[1]+1)] = numpy.zeros(curr[1] - curr[0] + 1)
                curr[2]=0
            else: #if curr[2]==0
                labels[curr[0] : (curr[1]+1)] = numpy.ones(curr[1] - curr[0] + 1)
                curr[2]=1

return labels

def mod_4hz(self, rate_wavsample):
        """Computes and returns the 4Hz modulation energy features for the given input wave file"""

# Set parameters
        wl = float(self.win_length_ms)
        ws = float(self.win_shift_ms)
        nf = self.n_filters
        f_min = float(self.f_min)
        f_max = float(self.f_max)
        pre = float(self.pre_emphasis_coef)

c = bob.ap.Spectrogram(float(rate_wavsample[0]), float(wl), float(ws), nf, float(f_min), float(f_max), float(pre))
        c.energy_filter=True
        c.log_filter=False
        c.energy_bands=True

sig =  rate_wavsample[1]
        energy_bands = c(sig)
        filtering_res = self.pass_band_filtering(energy_bands, rate_wavsample[0])
        mod_4hz = self.modulation_4hz(filtering_res, rate_wavsample)
        mod_4hz = self.averaging(mod_4hz)
        e = bob.ap.Energy(float(rate_wavsample[0]), float(wl), float(ws))
        energy_array = e(rate_wavsample[1])
        labels = self._voice_activity_detection(energy_array, mod_4hz)
        labels = self.smoothing(labels,self.smoothing_window) # discard isolated speech less than 100ms
        return labels, energy_array, mod_4hz

def __call__(self, input_signal, annotations=None):
        """labels speech (1) and non-speech (0) parts of the given input wave file using 4Hz modulation energy and energy
            Input parameter:
               * input_signal[0] --> rate
               * input_signal[1] --> signal
        """
        [labels, energy_array, mod_4hz] = self.mod_4hz(input_signal)
        rate    =  input_signal[0]
        data = input_signal[1]
        return rate, data, labels

class Algorithm:
    def __init__(self):
        self.win_length_ms = 20
        self.win_shift_ms = 10
        self.rate = 16000

def setup(self, parameters):
        self.rate = parameters.get('rate', self.rate)
        self.win_length_ms = parameters.get('win_length_ms', self.win_length_ms)
        self.win_shift_ms = parameters.get('win_shift_ms', self.win_shift_ms)
        self.preprocessor = Mod_4Hz(win_length_ms=self.win_length_ms, win_shift_ms=self.win_shift_ms)
        return True

def process(self, inputs, outputs):
        float_wav = inputs["speech"].data.value.astype('float64')

if float_wav is None or not float_wav.size:
            labels = numpy.zeros(2, dtype=numpy.int8)
        else:
            [labels, energies, mod_4hz] = self.preprocessor.mod_4hz([self.rate, float_wav])

outputs["labels"].write({
            'value': labels
        })

return True

The code for this algorithm in Python
The ruler at 80 columns indicate suggested POSIX line breaks (for readability).
The editor will automatically enlarge to accomodate the entirety of your input
Use keyboard shortcuts for search/replace and faster editing. For example, use Ctrl-F (PC) or Cmd-F (Mac) to search through this box

4Hz modulation of energy voice activity detection (VAD) with carefully tuned thresholds.

Experiments

Attestation:

Privacy:

Status:

Name	Databases/Protocols	Analyzers
pkorshunov/pkorshunov/isv-asv-pad-fusion-complete/1/asv_isv-pad_lbp_hist_ratios_lr-fusion_lr-pa_aligned	avspoof/2@physicalaccess_verify_train,avspoof/2@physicalaccess_verification,avspoof/2@physicalaccess_verification_spoof,avspoof/2@physicalaccess_verify_train_spoof,avspoof/2@physicalaccess_antispoofing	pkorshunov/spoof-score-fusion-roc_hist/1
pkorshunov/pkorshunov/speech-pad-simple/1/speech-pad_lbp_hist_ratios_lr-pa_aligned	avspoof/2@physicalaccess_antispoofing	pkorshunov/simple_antispoofing_analyzer/4
pkorshunov/pkorshunov/isv-asv-pad-fusion-complete/1/asv_isv-pad_gmm-fusion_lr-pa	avspoof/2@physicalaccess_verify_train,avspoof/2@physicalaccess_verification,avspoof/2@physicalaccess_verification_spoof,avspoof/2@physicalaccess_verify_train_spoof,avspoof/2@physicalaccess_antispoofing	pkorshunov/spoof-score-fusion-roc_hist/1
pkorshunov/pkorshunov/speech-pad-simple/1/speech-pad_gmm-pa	avspoof/2@physicalaccess_antispoofing	pkorshunov/simple_antispoofing_analyzer/4
pkorshunov/pkorshunov/isv-speaker-verification-spoof/1/isv-speaker-verification-spoof-pa	avspoof/2@physicalaccess_verification_spoof,avspoof/2@physicalaccess_verification	pkorshunov/eerhter_postperf_iso_spoof/1
pkorshunov/pkorshunov/isv-speaker-verification/1/isv-speaker-verification-licit	avspoof/2@physicalaccess_verification	pkorshunov/eerhter_postperf_iso/1
pkorshunov/pkorshunov/speech-antispoofing-baseline/1/btas2016-baseline-pa	avspoof/1@physicalaccess_antispoofing	pkorshunov/simple_antispoofing_analyzer/2

This table shows the number of times this algorithm has been successfully run using the given environment. Note this does not provide sufficient information to evaluate if the algorithm will run when submitted to different conditions.

algorithms pkorshunov vad_4hz 1

Endpoint Groups 1

Group: main

Parameters 3

Experiments

algorithms

pkorshunov

vad_4hz

1