Numpy Models and Tools

Hyperion provides several models and feature extractors based on numpy.

Feature Extraction and Voice Activity Detection

Feature Extraction Classes

Copyright 2018 Johns Hopkins University (Author: Jesus Villalba) Apache 2.0 (http://www.apache.org/licenses/LICENSE-2.0)

class hyperion.feats.mfcc.MFCCSteps(value)

Steps in the MFCC pipeline

WAVE = 0

FFT = 1

SPEC = 2

LOG_SPEC = 3

LOGFB = 4

MFCC = 5

class hyperion.feats.mfcc.MFCC(sample_frequency=16000, frame_length=25, frame_shift=10, fft_length=512, remove_dc_offset=True, preemphasis_coeff=0.97, window_type='povey', use_fft2=True, dither=1, fb_type='mel_kaldi', low_freq=20, high_freq=0, num_filters=23, norm_filters=False, num_ceps=13, snip_edges=True, energy_floor=0, raw_energy=True, use_energy=True, cepstral_lifter=22, input_step='wave', output_step='mfcc')

Compute MFCC features.

sample_frequency

Waveform data sample frequency (must match the waveform file, if specified there) (default = 16000)

frame_length

Frame length in milliseconds (default = 25)

frame_shift

Frame shift in milliseconds (default = 10)

fft_length

Length of FFT (default = 512)

remove_dc_offset

Subtract mean from waveform on each frame (default = True)

preemphasis_coeff

Coefficient for use in signal preemphasis (default = 0.97)

window_type

Type of window (“hamming”|”hanning”|”povey”|”rectangular”|”blackmann”) (default = ‘povey’)

use_fft2

If true, it uses |X(f)|^2, if false, it uses |X(f)|, (default = True)

dither

Dithering constant (0.0 means no dither) (default = 1)

fb_type

Filter-bank type: mel_kaldi, mel_etsi, mel_librosa, mel_librosa_htk, linear (default = ‘mel_kaldi’)

low_freq

Low cutoff frequency for mel bins (default = 20)

high_freq

High cutoff frequency for mel bins (if < 0, offset from Nyquist) (default = 0)

num_filters

Number of triangular mel-frequency bins (default = 23)

norm_filters

Normalize filters coeff to sum up to 1, if librosa it uses stanley norm (default = False)

num_ceps

Number of cepstra in MFCC computation (including C0) (default = 13)

snip_edges

If true, end effects will be handled by outputting only frames that completely fit in the file, and the number of frames depends on the frame-length. If false, the number of frames depends only on the frame-shift, and we reflect the data at the ends. (default = True)

energy_floor

Floor on energy (absolute, not relative) in MFCC computation (default = 0)

raw_energy

If true, compute energy before preemphasis and windowing (default = True)

use_energy

Use energy (not C0) in MFCC computation (default = True)

cepstral_lifter

Constant that controls scaling of MFCCs (default = 22)

input_step

It can continue computation from any step: wav, fft, spec, logfb (default = ‘wav’)

output_step

It can return intermediate result: fft, spec, logfb, mfcc (default = ‘mfcc’)

__init__(sample_frequency=16000, frame_length=25, frame_shift=10, fft_length=512, remove_dc_offset=True, preemphasis_coeff=0.97, window_type='povey', use_fft2=True, dither=1, fb_type='mel_kaldi', low_freq=20, high_freq=0, num_filters=23, norm_filters=False, num_ceps=13, snip_edges=True, energy_floor=0, raw_energy=True, use_energy=True, cepstral_lifter=22, input_step='wave', output_step='mfcc')

reset()

Resets the internal states of the filters

static make_lifter(N, Q)

Makes the liftering function

Parameters

• N – Number of cepstral coefficients.

• Q – Liftering parameter

Returns

Liftering vector.

compute_raw_logE(x)

Computes log-energy before preemphasis filter

Parameters

x – wave signal

Returns

Log-energy

compute(x, return_fft=False, return_spec=False, return_logfb=False)

Evaluates the MFCC pipeline.

Parameters

• x – Wave, stft, spectrogram or log-filter-bank depending on input_step.

• return_fft – If true, it also returns short-time fft.

• return_spec – If true, it also returns short-time magnitude spectrogram.

• return_logfb – If true, it also returns log-filter-bank.

Returns

Stfft, spectrogram, log-filter-bank or MFCC depending on output_step.

static filter_args(**kwargs)

Filters MFCC args from arguments dictionary.

Parameters

kwargs – Arguments dictionary.

Returns

Dictionary with MFCC options.

static add_class_args(parser, prefix=None)

Adds MFCC options to parser.

Parameters

• parser – Arguments parser

• prefix – Options prefix.

static add_argparse_args(parser, prefix=None)

Adds MFCC options to parser.

Parameters

• parser – Arguments parser

• prefix – Options prefix.

class hyperion.feats.filter_banks.FilterBankFactory

static create(filter_bank_type, num_filters, fft_length, fs, low_freq, high_freq, norm_filters)

static lin2mel(x)

static mel2lin(x)

static make_mel_kaldi(num_filters, fft_length, fs, low_freq, high_freq)

static make_mel_etsi(num_filters, fft_length, fs, low_freq, high_freq)

static make_linear(num_filters, fft_length, fs, low_freq, high_freq)

static make_mel_librosa(num_filters, fft_length, fs, low_freq, high_freq, htk=False, norm_filters=False)

static add_class_args(parser, prefix=None)

static add_argparse_args(parser, prefix=None)

class hyperion.feats.feature_windows.FeatureWindowFactory

static create(window_type, N, sym=False)

static add_class_args(parser, prefix=None)

static add_argparse_args(parser, prefix=None)

Feature Normalization Classes

Copyright 2019 Johns Hopkins University (Author: Jesus Villalba) Apache 2.0 (http://www.apache.org/licenses/LICENSE-2.0)

class hyperion.feats.feature_normalization.MeanVarianceNorm(norm_mean=True, norm_var=False, left_context=None, right_context=None)

Class to perform mean and variance normalization

norm_mean

normalize mean

norm_var

normalize variance

left_context

past context of the sliding window, if None all past frames.

right_context

future context of the sliding window, if None all future frames.

If left_context==right_context==None, it will apply global mean/variance normalization.

__init__(norm_mean=True, norm_var=False, left_context=None, right_context=None)

normalize(x)

normalize_global(x)

normalize_conv(x)

Normalize featurex in x

Uses convolution operator

Parameters

x – Input feature matrix.

Returns

Normalized feature matrix.

normalize_cumsum(x)

Normalize featurex in x

Uses cumsum

Parameters

x – Input feature matrix.

Returns

Normalized feature matrix.

normalize_slow(x)

static filter_args(**kwargs)

Filters ST-CMVN args from arguments dictionary.

Parameters

• prefix – Options prefix.

• kwargs – Arguments dictionary.

Returns

Dictionary with ST-CMVN options.

static add_class_args(parser, prefix=None)

Adds ST-CMVN options to parser.

Parameters

• parser – Arguments parser

• prefix – Options prefix.

static add_argparse_args(parser, prefix=None)

Adds ST-CMVN options to parser.

Parameters

• parser – Arguments parser

• prefix – Options prefix.

Voice Activity Detection Classes

class hyperion.feats.energy_vad.EnergyVAD(sample_frequency=16000, frame_length=25, frame_shift=10, dither=1, snip_edges=True, vad_energy_mean_scale=0.5, vad_energy_threshold=5, vad_frames_context=0, vad_proportion_threshold=0.6)

Compute VAD based on Kaldi Energy VAD method.

sample_frequency

Waveform data sample frequency (must match the waveform file, if specified there) (default = 16000)

frame_length

Frame length in milliseconds (default = 25)

frame_shift

Frame shift in milliseconds (default = 10)

dither

Dithering constant (0.0 means no dither) (default = 1)

snip_edges

If true, end effects will be handled by outputting only frames that completely fit in the file, and the number of frames depends on the frame-length. If false, the number of frames depends only on the frame-shift, and we reflect the data at the ends. (default = True)

vad_energy_mean_scale

If this is set to s, to get the actual threshold we let m be the mean log-energy of the file, and use s*m + vad-energy-threshold (float, default = 0.5)

vad_energy_threshold

Constant term in energy threshold for MFCC0 for VAD (also see –vad-energy-mean-scale) (float, default = 5)

vad_frames_context

Number of frames of context on each side of central frame, in window for which energy is monitored (int, default = 0)

vad_proportion_threshold

Parameter controlling the proportion of frames within the window that need to have more energy than the threshold (float, default = 0.6)

__init__(sample_frequency=16000, frame_length=25, frame_shift=10, dither=1, snip_edges=True, vad_energy_mean_scale=0.5, vad_energy_threshold=5, vad_frames_context=0, vad_proportion_threshold=0.6)

reset()

Resets the internal states of the filters

compute(x, return_loge=False)

Evaluates the VAD.

Parameters

• x – Wave

• return_loge – If true, it also returns the log-energy.

Returns

Binary VAD

static filter_args(**kwargs)

Filters VAD args from arguments dictionary.

Parameters

kwargs – Arguments dictionary.

Returns

Dictionary with VAD options.

static add_class_args(parser, prefix=None)

Adds VAD options to parser.

Parameters

• parser – Arguments parser

• prefix – Options prefix.

static add_argparse_args(parser, prefix=None)

Adds VAD options to parser.

Parameters

• parser – Arguments parser

• prefix – Options prefix.

Feature Extraction Functions

Copyright 2018 Johns Hopkins University (Author: Jesus Villalba) Apache 2.0 (http://www.apache.org/licenses/LICENSE-2.0)

hyperion.feats.stft.stft(x, frame_length, frame_shift, fft_length, window=None)

hyperion.feats.stft.istft(X, frame_length, frame_shift, window=None)

hyperion.feats.stft.strft(x, frame_length, frame_shift, fft_length, window=None)

hyperion.feats.stft.istrft(X, frame_length, frame_shift, window=None)

hyperion.feats.stft.st_logE(x, frame_length, frame_shift)

Computes log-energy before preemphasis filter

Parameters

x – wave signal

Returns

Log-energy

Speech Augmentation

Combined Speech Augmentation Class

class hyperion.augment.speech_augment.SpeechAugment(speed_aug=None, reverb_aug=None, noise_aug=None)

Class to add noise and reverberation on-the-fly when trianing nnets.

speed_aug

SpeedAugment object

reverb_aug

ReverbAugment object

noise_aug

NoiseAugment object

__init__(speed_aug=None, reverb_aug=None, noise_aug=None)

classmethod create(cfg, random_seed=112358, rng=None)

property max_reverb_context

forward(x)

Noise Augmentation Classes

class hyperion.augment.noise_augment.NoiseAugment(noise_prob, noise_types, random_seed=112358, rng=None)

Class to augment speech with additive noise from multiple types,

e.g., music, babble, … It will randomly choose which noise type to add.

noise_prob

probability of adding noise

noise_types

dictionary of options with one entry per noise-type, Each entry is also a dictiory with the following entries: weight, max_snr, min_snr, noise_path. The weight parameter is proportional to how often we want to sample a given noise type.

rng

Random number generator returned by np.random.RandomState (optional)

__init__(noise_prob, noise_types, random_seed=112358, rng=None)

classmethod create(cfg, random_seed=112358, rng=None)

Creates a NoiseAugment object from options dictionary or YAML file.

Parameters

• cfg – YAML file path or dictionary with noise options.

• rng – Random number generator returned by np.random.RandomState (optional)

Returns

NoiseAugment object

forward(x)

class hyperion.augment.noise_augment.SingleNoiseAugment(noise_type, noise_path, min_snr, max_snr, random_seed=112358, rng=None)

Class to augment speech with additive noise of a single type,

e.g., music, babble, …

noise_type

string label indicating the noise type.

noise_path

path to Kaldi style wav.scp file indicating the path to the noise wav files.

min_snr

mininimum SNR(dB) to sample from.

max_snr

maximum SNR(dB) to sample from.

rng

Random number generator returned by np.random.RandomState (optional)

__init__(noise_type, noise_path, min_snr, max_snr, random_seed=112358, rng=None)

static snr(x, n)

forward(x)

Reverberation Augmentation Classes

class hyperion.augment.reverb_augment.ReverbAugment(reverb_prob, rir_types, max_reverb_context=0, random_seed=112358, rng=None)

Class to augment speech with reverberation with RIRS from multiple types,

e.g., small room, medium room, large room. It will randomly choose which RIR type to add.

reverb_prob

probability of adding reverberation

rir_types

dictionary of options with one entry per RIR-type, Each entry is also a dictiory with the following entries: weight, rir_norm, comp_delay, rir_path. The weight parameter is proportional to how often we want to sample a given RIR type.

max_reverb_context

number of samples required as left context for the convolution operation.

rng

Random number generator returned by np.random.RandomState (optional)

__init__(reverb_prob, rir_types, max_reverb_context=0, random_seed=112358, rng=None)

classmethod create(cfg, random_seed=112358, rng=None)

Creates a ReverbAugment object from options dictionary or YAML file.

Parameters

• cfg – YAML file path or dictionary with reverb options.

• rng – Random number generator returned by np.random.RandomState (optional)

Returns

ReverbAugment object

static sdr(x, y, scale, delay)

forward(x)

class hyperion.augment.reverb_augment.SingleReverbAugment(rir_type, rir_path, rir_norm=None, comp_delay=True, preload_rirs=True, random_seed=112358, rng=None)

Class to augment speech with reverberation using RIR from a

single type, e.g., small room, medium room, large room

rir_type

string label indicating the RIR type.

rir_path

Kaldi style rspecifier to Ark or H5 file containing RIRs

rir_norm

RIR normalization method between None, ‘max’ or ‘energy’

comp_delay

compensate the delay introduced by the RIR if any, this delay will happen if the maximum of the RIR is not in its first sample.

preload_rirs

if True all RIRS are loaded into RAM

rng

Random number generator returned by np.random.RandomState (optional)

__init__(rir_type, rir_path, rir_norm=None, comp_delay=True, preload_rirs=True, random_seed=112358, rng=None)

static sdr(x, y, scale, delay)

forward(x)

Speed Augmentation Classes

class hyperion.augment.speed_augment.SpeedAugment(speed_prob, speed_ratios=[0.9, 1.1], keep_length=False, random_seed=112358, rng=None)

Class to augment speech with speed perturbation

speed_prob

probability of applying speed perturbation

speed_ratios

list of speed pertubation ratios

keep_length

applies padding or cropping to keep the lenght of the signal

random_seed

random seed for random number generator

rng

Random number generator returned by np.random.RandomState (optional)

__init__(speed_prob, speed_ratios=[0.9, 1.1], keep_length=False, random_seed=112358, rng=None)

classmethod create(cfg, random_seed=112358, rng=None)

Creates a SpeedAugment object from options dictionary or YAML file.

Parameters

• cfg – YAML file path or dictionary with noise options.

• rng – Random number generator returned by np.random.RandomState (optional)

Returns

NoiseAugment object

forward(x)

Hyperion Numpy Models

All numpy ML models in Hyperion derive from the same base class

class hyperion.hyp_model.HypModel(name=None, **kwargs)

__init__(name=None, **kwargs)

copy()

property is_init

init_to_false()

abstract initialize()

abstract fit(x, sample_weights=None, x_val=None, sample_weights_val=None)

abstract fit_generator(x, x_val=None)

abstract save(file_path)

abstract save_params(f)

classmethod load_config(file_path)

classmethod load(file_path)

classmethod load_params(f, config)

abstract get_config()

to_json(**kwargs)

static load_config_from_json(json_str)

class hyperion.model_loader.ModelLoader

static get_object()

static load(file_path)

Probability Density Functions

These are classes that define different probability density functions like GMMs and PLDA

Core PDF Classes

class hyperion.pdfs.core.pdf.PDF(x_dim=1, **kwargs)

__init__(x_dim=1, **kwargs)

get_config()

abstract log_prob(x)

eval_llk(x)

abstract sample(num_samples)

generate(num_samples, **kwargs)

copy()

abstract fit(x, sample_weights=None, x_val=None, sample_weights_val=None)

abstract fit_generator(x, x_val=None)

init_to_false()

abstract initialize()

property is_init

classmethod load(file_path)

classmethod load_config(file_path)

static load_config_from_json(json_str)

classmethod load_params(f, config)

abstract save(file_path)

abstract save_params(f)

to_json(**kwargs)

class hyperion.pdfs.core.exp_family.ExpFamily(eta=None, **kwargs)

__init__(eta=None, **kwargs)

property is_init

fit(x, sample_weight=None, x_val=None, sample_weight_val=None, batch_size=None)

log_h(x)

accum_log_h(x, sample_weight=None)

compute_suff_stats(x)

accum_suff_stats(x, u_x=None, sample_weight=None, batch_size=None)

add_suff_stats(N, u_x)

Estep(x, u_x=None, sample_weight=None, batch_size=None)

abstract Mstep(stats)

elbo(x, u_x=None, N=1, log_h=None, sample_weight=None, batch_size=None)

log_prob(x, u_x=None, method='nat')

log_prob_nat(x, u_x=None)

static compute_A_nat(eta)

static compute_A_std(params)

static compute_eta(param)

static compute_std(eta)

abstract validate()

copy()

eval_llk(x)

abstract fit_generator(x, x_val=None)

generate(num_samples, **kwargs)

get_config()

init_to_false()

abstract initialize()

classmethod load(file_path)

classmethod load_config(file_path)

static load_config_from_json(json_str)

classmethod load_params(f, config)

abstract sample(num_samples)

abstract save(file_path)

abstract save_params(f)

to_json(**kwargs)

class hyperion.pdfs.core.normal_diag_cov.NormalDiagCov(mu=None, Lambda=None, var_floor=1e-05, update_mu=True, update_Lambda=True, **kwargs)

__init__(mu=None, Lambda=None, var_floor=1e-05, update_mu=True, update_Lambda=True, **kwargs)

property logLambda

property cholLambda

property Sigma

initialize()

stack_suff_stats(F, S=None)

unstack_suff_stats(stats)

norm_suff_stats(N, u_x=None, return_order2=False)

Mstep(N, u_x)

log_prob_std(x)

log_cdf(x)

sample(num_samples, rng=None, seed=1024)

get_config()

save_params(f)

classmethod load_params(f, config)

validate()

static compute_eta(mu, Lambda)

static compute_std(eta)

static compute_A_nat(eta)

static compute_A_std(mu, Lambda)

static compute_suff_stats(x)

plot1D(feat_idx=0, num_sigmas=2, num_pts=100, **kwargs)

plot2D(feat_idx=[0, 1], num_sigmas=2, num_pts=100, **kwargs)

plot3D(feat_idx=[0, 1], num_sigmas=2, num_pts=100, **kwargs)

plot3D_ellipsoid(feat_idx=[0, 1, 2], num_sigmas=2, num_pts=100, **kwargs)

Estep(x, u_x=None, sample_weight=None, batch_size=None)

accum_log_h(x, sample_weight=None)

accum_suff_stats(x, u_x=None, sample_weight=None, batch_size=None)

add_suff_stats(N, u_x)

copy()

elbo(x, u_x=None, N=1, log_h=None, sample_weight=None, batch_size=None)

eval_llk(x)

fit(x, sample_weight=None, x_val=None, sample_weight_val=None, batch_size=None)

abstract fit_generator(x, x_val=None)

generate(num_samples, **kwargs)

init_to_false()

property is_init

classmethod load(file_path)

classmethod load_config(file_path)

static load_config_from_json(json_str)

log_h(x)

log_prob(x, u_x=None, method='nat')

log_prob_nat(x, u_x=None)

abstract save(file_path)

to_json(**kwargs)

PLDA Classes

class hyperion.pdfs.plda.plda_base.PLDABase(y_dim=None, mu=None, update_mu=True, **kwargs)

__init__(y_dim=None, mu=None, update_mu=True, **kwargs)

abstract initialize(D)

abstract compute_py_g_x(D)

fit(x, class_ids=None, ptheta=None, sample_weight=None, x_val=None, class_ids_val=None, ptheta_val=None, sample_weight_val=None, epochs=20, ml_md='ml+md', md_epochs=None)

abstract Estep(x)

abstract MstepML(x)

abstract MstepMD(x)

abstract llr_1vs1(x1, x2)

abstract llr_NvsM_book(D1, D2)

fit_adapt_weighted_avg_model(x, class_ids=None, ptheta=None, sample_weight=None, x_val=None, class_ids_val=None, ptheta_val=None, sample_weight_val=None, epochs=20, ml_md='ml+md', md_epochs=None, plda0=None, w_mu=1, w_B=0.5, w_W=0.5)

fit_adapt(x, class_ids=None, ptheta=None, sample_weight=None, x0=None, class_ids0=None, ptheta0=None, sample_weight0=None, x_val=None, class_ids_val=None, ptheta_val=None, sample_weight_val=None, epochs=20, ml_md='ml+md', md_epochs=None)

static compute_stats_soft(x, p_theta, sample_weight=None, scal_factor=None)

static compute_stats_hard(x, class_ids, sample_weight=None, scale_factor=None)

static compute_stats_hard_v0(x, class_ids, sample_weight=None, scal_factor=None)

static center_stats(D, mu)

llr_NvsM(x1, x2, ids1=None, ids2=None, method='vavg-lnorm')

llr_NvsM_vavg(D1, D2, do_lnorm=True)

llr_NvsM_savg(x1, ids1, x2, ids2)

llr_Nvs1(x1, x2, ids1=None, method='vavg-lnorm')

llr_Nvs1_vavg(D1, x2, do_lnorm=True)

llr_Nvs1_savg(x1, ids1, x2)

abstract sample(num_classes, num_samples_per_class, rng=None, seed=1024)

get_config()

weigthed_avg_params(mu, w_mu)

abstract weigthed_avg_model(plda)

copy()

eval_llk(x)

abstract fit_generator(x, x_val=None)

generate(num_samples, **kwargs)

init_to_false()

property is_init

classmethod load(file_path)

classmethod load_config(file_path)

static load_config_from_json(json_str)

classmethod load_params(f, config)

abstract log_prob(x)

abstract save(file_path)

abstract save_params(f)

to_json(**kwargs)

class hyperion.pdfs.plda.frplda.FRPLDA(mu=None, B=None, W=None, fullcov_W=True, update_mu=True, update_B=True, update_W=True, **kwargs)

__init__(mu=None, B=None, W=None, fullcov_W=True, update_mu=True, update_B=True, update_W=True, **kwargs)

validate()

property is_init

initialize(D)

compute_py_g_x(D, return_cov=False, return_logpy_0=False, return_acc=False)

Estep(D)

elbo(stats)

MstepML(stats)

MstepMD(stats)

get_config()

save_params(f)

classmethod load_params(f, config)

llr_1vs1(x1, x2)

llr_NvsM_book(D1, D2)

sample(num_classes, num_samples_per_class, rng=None, seed=1024, return_y=False)

weighted_avg_params(mu, B, W, w_mu, w_B, w_W)

weighted_avg_model(plda, w_mu, w_B, w_W)

static center_stats(D, mu)

static compute_stats_hard(x, class_ids, sample_weight=None, scale_factor=None)

static compute_stats_hard_v0(x, class_ids, sample_weight=None, scal_factor=None)

static compute_stats_soft(x, p_theta, sample_weight=None, scal_factor=None)

copy()

eval_llk(x)

fit(x, class_ids=None, ptheta=None, sample_weight=None, x_val=None, class_ids_val=None, ptheta_val=None, sample_weight_val=None, epochs=20, ml_md='ml+md', md_epochs=None)

fit_adapt(x, class_ids=None, ptheta=None, sample_weight=None, x0=None, class_ids0=None, ptheta0=None, sample_weight0=None, x_val=None, class_ids_val=None, ptheta_val=None, sample_weight_val=None, epochs=20, ml_md='ml+md', md_epochs=None)

fit_adapt_weighted_avg_model(x, class_ids=None, ptheta=None, sample_weight=None, x_val=None, class_ids_val=None, ptheta_val=None, sample_weight_val=None, epochs=20, ml_md='ml+md', md_epochs=None, plda0=None, w_mu=1, w_B=0.5, w_W=0.5)

abstract fit_generator(x, x_val=None)

generate(num_samples, **kwargs)

init_to_false()

llr_Nvs1(x1, x2, ids1=None, method='vavg-lnorm')

llr_Nvs1_savg(x1, ids1, x2)

llr_Nvs1_vavg(D1, x2, do_lnorm=True)

llr_NvsM(x1, x2, ids1=None, ids2=None, method='vavg-lnorm')

llr_NvsM_savg(x1, ids1, x2, ids2)

llr_NvsM_vavg(D1, D2, do_lnorm=True)

classmethod load(file_path)

classmethod load_config(file_path)

static load_config_from_json(json_str)

abstract log_prob(x)

abstract save(file_path)

to_json(**kwargs)

abstract weigthed_avg_model(plda)

weigthed_avg_params(mu, w_mu)

class hyperion.pdfs.plda.splda.SPLDA(y_dim=None, mu=None, V=None, W=None, fullcov_W=True, update_mu=True, update_V=True, update_W=True, **kwargs)

__init__(y_dim=None, mu=None, V=None, W=None, fullcov_W=True, update_mu=True, update_V=True, update_W=True, **kwargs)

validate()

property is_init

initialize(D)

compute_py_g_x(D, return_cov=False, return_logpy_0=False, return_acc=False)

Estep(D)

elbo(stats)

MstepML(stats)

MstepMD(stats)

get_config()

save_params(f)

classmethod load_params(f, config)

log_probx_g_y(x, y)

llr_1vs1(x1, x2)

llr_NvsM_book(D1, D2)

sample(num_classes, num_samples_per_class, rng=None, seed=1024)

weighted_avg_params(mu, V, W, w_mu, w_B, w_W)

weighted_avg_model(plda, w_mu, w_B, w_W)

project(T, delta_mu=None)

static center_stats(D, mu)

static compute_stats_hard(x, class_ids, sample_weight=None, scale_factor=None)

static compute_stats_hard_v0(x, class_ids, sample_weight=None, scal_factor=None)

static compute_stats_soft(x, p_theta, sample_weight=None, scal_factor=None)

copy()

eval_llk(x)

fit(x, class_ids=None, ptheta=None, sample_weight=None, x_val=None, class_ids_val=None, ptheta_val=None, sample_weight_val=None, epochs=20, ml_md='ml+md', md_epochs=None)

fit_adapt(x, class_ids=None, ptheta=None, sample_weight=None, x0=None, class_ids0=None, ptheta0=None, sample_weight0=None, x_val=None, class_ids_val=None, ptheta_val=None, sample_weight_val=None, epochs=20, ml_md='ml+md', md_epochs=None)

fit_adapt_weighted_avg_model(x, class_ids=None, ptheta=None, sample_weight=None, x_val=None, class_ids_val=None, ptheta_val=None, sample_weight_val=None, epochs=20, ml_md='ml+md', md_epochs=None, plda0=None, w_mu=1, w_B=0.5, w_W=0.5)

abstract fit_generator(x, x_val=None)

generate(num_samples, **kwargs)

init_to_false()

llr_Nvs1(x1, x2, ids1=None, method='vavg-lnorm')

llr_Nvs1_savg(x1, ids1, x2)

llr_Nvs1_vavg(D1, x2, do_lnorm=True)

llr_NvsM(x1, x2, ids1=None, ids2=None, method='vavg-lnorm')

llr_NvsM_savg(x1, ids1, x2, ids2)

llr_NvsM_vavg(D1, D2, do_lnorm=True)

classmethod load(file_path)

classmethod load_config(file_path)

static load_config_from_json(json_str)

abstract log_prob(x)

abstract save(file_path)

to_json(**kwargs)

abstract weigthed_avg_model(plda)

weigthed_avg_params(mu, w_mu)

class hyperion.pdfs.plda.plda.PLDA(y_dim=None, z_dim=None, mu=None, V=None, U=None, D=None, floor_iD=1e-05, update_mu=True, update_V=True, update_U=True, update_D=True, **kwargs)

__init__(y_dim=None, z_dim=None, mu=None, V=None, U=None, D=None, floor_iD=1e-05, update_mu=True, update_V=True, update_U=True, update_D=True, **kwargs)

validate()

property is_init

compute_aux()

initialize(D)

compute_py_g_x(D, return_cov=False, return_logpy_0=False, return_acc=False)

Estep(D)

elbo(stats)

MstepML(stats)

MstepMD(stats)

get_config()

save_params(f)

classmethod load_params(f, config)

log_probx_g_y(x, y)

log_probx_g_yz(x, y, z)

llr_1vs1(x1, x2)

llr_NvsM_book(D1, D2)

sample(num_classes, num_samples_per_class, rng=None, seed=1024)

weighted_avg_params(mu, V, U, D, w_mu, w_B, w_W)

static center_stats(D, mu)

static compute_stats_hard(x, class_ids, sample_weight=None, scale_factor=None)

static compute_stats_hard_v0(x, class_ids, sample_weight=None, scal_factor=None)

static compute_stats_soft(x, p_theta, sample_weight=None, scal_factor=None)

copy()

eval_llk(x)

fit(x, class_ids=None, ptheta=None, sample_weight=None, x_val=None, class_ids_val=None, ptheta_val=None, sample_weight_val=None, epochs=20, ml_md='ml+md', md_epochs=None)

fit_adapt(x, class_ids=None, ptheta=None, sample_weight=None, x0=None, class_ids0=None, ptheta0=None, sample_weight0=None, x_val=None, class_ids_val=None, ptheta_val=None, sample_weight_val=None, epochs=20, ml_md='ml+md', md_epochs=None)

fit_adapt_weighted_avg_model(x, class_ids=None, ptheta=None, sample_weight=None, x_val=None, class_ids_val=None, ptheta_val=None, sample_weight_val=None, epochs=20, ml_md='ml+md', md_epochs=None, plda0=None, w_mu=1, w_B=0.5, w_W=0.5)

abstract fit_generator(x, x_val=None)

generate(num_samples, **kwargs)

init_to_false()

llr_Nvs1(x1, x2, ids1=None, method='vavg-lnorm')

llr_Nvs1_savg(x1, ids1, x2)

llr_Nvs1_vavg(D1, x2, do_lnorm=True)

llr_NvsM(x1, x2, ids1=None, ids2=None, method='vavg-lnorm')

llr_NvsM_savg(x1, ids1, x2, ids2)

llr_NvsM_vavg(D1, D2, do_lnorm=True)

classmethod load(file_path)

classmethod load_config(file_path)

static load_config_from_json(json_str)

abstract log_prob(x)

abstract save(file_path)

to_json(**kwargs)

weighted_avg_model(plda, w_mu, w_B, w_W)

abstract weigthed_avg_model(plda)

weigthed_avg_params(mu, w_mu)

Mixture Models

class hyperion.pdfs.mixtures.exp_family_mixture.ExpFamilyMixture(num_comp=1, pi=None, eta=None, min_N=0, update_pi=True, **kwargs)

__init__(num_comp=1, pi=None, eta=None, min_N=0, update_pi=True, **kwargs)

property is_init

property log_pi

fit(x, sample_weight=None, x_val=None, sample_weight_val=None, epochs=10, batch_size=None)

fit_generator(generator, train_steps, epochs=10, val_data=None, val_steps=0, max_queue_size=10, workers=1, use_multiprocessing=False)

log_h(x)

accum_log_h(x, sample_weight=None)

compute_log_pz(x, u_x=None, mode='nat')

compute_pz(x, u_x=None, mode='nat')

compute_pz_nat(x, u_x=None)

compute_pz_std(x)

compute_suff_stats(x)

accum_suff_stats(x, u_x=None, sample_weight=None, batch_size=None)

accum_suff_stats_segments(x, segments, u_x=None, sample_weight=None, batch_size=None)

accum_suff_stats_segments_prob(x, prob, u_x=None, sample_weight=None, batch_size=None)

accum_suff_stats_sorttime(x, frame_length, frame_shift, u_x=None, sample_weight=None, batch_size=None)

Estep(x, u_x=None, sample_weight=None, batch_size=None)

Estep_generator(generator, num_steps, return_log_h, max_queue_size=10, workers=1, use_multiprocessin=False)

sum_suff_stats(N, u_x)

abstract Mstep(stats)

elbo(x, u_x=None, N=1, log_h=None, sample_weight=None, batch_size=None)

log_prob(x, u_x=None, mode='nat')

log_prob_nat(x, u_x=None)

abstract log_prob_std(x)

log_prob_nbest(x, u_x=None, mode='nat', nbest_mode='master', nbest=1)

log_prob_nbest_nat(x, u_x=None, nbest_mode='master', nbest=1)

abstract log_prob_nbest_std(x, nbest_mode='master', nbest=1)

get_config()

static tuple2data(data)

static compute_A_nat(eta)

static compute_A_std(params)

static compute_eta(param)

static compute_std(eta)

copy()

eval_llk(x)

generate(num_samples, **kwargs)

init_to_false()

abstract initialize()

classmethod load(file_path)

classmethod load_config(file_path)

static load_config_from_json(json_str)

classmethod load_params(f, config)

abstract sample(num_samples)

abstract save(file_path)

abstract save_params(f)

to_json(**kwargs)

class hyperion.pdfs.mixtures.gmm.GMM(mu=None, Lambda=None, var_floor=0.001, update_mu=True, update_Lambda=True, **kwargs)

__init__(mu=None, Lambda=None, var_floor=0.001, update_mu=True, update_Lambda=True, **kwargs)

compute_Lambda_aux()

property logLambda

property cholLambda

property Sigma

initialize(x=None)

stack_suff_stats(F, S=None)

unstack_suff_stats(stats)

norm_suff_stats(N, u_x, return_order2=False)

Mstep(N, u_x)

split_comp(K=2)

log_prob_std(x)

sample(num_samples, rng=None, seed=1024)

get_config()

save_params(f)

classmethod load_params(f, config)

classmethod load_from_kaldi(file_path)

validate()

static compute_eta(mu, Lambda)

static compute_std(eta)

static compute_A_nat(eta)

static compute_A_std(mu, Lambda)

static compute_suff_stats(x)

plot1D(feat_idx=0, num_sigmas=2, num_pts=100, **kwargs)

plot2D(feat_idx=[0, 1], num_sigmas=2, num_pts=100, **kwargs)

plot3D(feat_idx=[0, 1], num_sigmas=2, num_pts=100, **kwargs)

plot3D_ellipsoid(feat_idx=[0, 1, 2], num_sigmas=2, num_pts=100, **kwargs)

Estep(x, u_x=None, sample_weight=None, batch_size=None)

Estep_generator(generator, num_steps, return_log_h, max_queue_size=10, workers=1, use_multiprocessin=False)

accum_log_h(x, sample_weight=None)

accum_suff_stats(x, u_x=None, sample_weight=None, batch_size=None)

accum_suff_stats_segments(x, segments, u_x=None, sample_weight=None, batch_size=None)

accum_suff_stats_segments_prob(x, prob, u_x=None, sample_weight=None, batch_size=None)

accum_suff_stats_sorttime(x, frame_length, frame_shift, u_x=None, sample_weight=None, batch_size=None)

compute_log_pz(x, u_x=None, mode='nat')

compute_pz(x, u_x=None, mode='nat')

compute_pz_nat(x, u_x=None)

compute_pz_std(x)

copy()

elbo(x, u_x=None, N=1, log_h=None, sample_weight=None, batch_size=None)

eval_llk(x)

fit(x, sample_weight=None, x_val=None, sample_weight_val=None, epochs=10, batch_size=None)

fit_generator(generator, train_steps, epochs=10, val_data=None, val_steps=0, max_queue_size=10, workers=1, use_multiprocessing=False)

generate(num_samples, **kwargs)

init_to_false()

property is_init

classmethod load(file_path)

classmethod load_config(file_path)

static load_config_from_json(json_str)

log_h(x)

property log_pi

log_prob(x, u_x=None, mode='nat')

log_prob_nat(x, u_x=None)

log_prob_nbest(x, u_x=None, mode='nat', nbest_mode='master', nbest=1)

log_prob_nbest_nat(x, u_x=None, nbest_mode='master', nbest=1)

abstract log_prob_nbest_std(x, nbest_mode='master', nbest=1)

abstract save(file_path)

sum_suff_stats(N, u_x)

to_json(**kwargs)

static tuple2data(data)

class hyperion.pdfs.mixtures.gmm_diag_cov.GMMDiagCov(mu=None, Lambda=None, var_floor=0.001, update_mu=True, update_Lambda=True, **kwargs)

__init__(mu=None, Lambda=None, var_floor=0.001, update_mu=True, update_Lambda=True, **kwargs)

property logLambda

property cholLambda

property Sigma

initialize(x=None)

stack_suff_stats(F, S=None)

unstack_suff_stats(stats)

norm_suff_stats(N, u_x, return_order2=False)

Mstep(N, u_x)

split_comp(K=2)

log_prob_std(x)

log_cdf(x)

sample(num_samples, rng=None, seed=1024)

get_config()

save_params(f)

classmethod load_params(f, config)

classmethod load_from_kaldi(file_path)

validate()

static compute_eta(mu, Lambda)

static compute_std(eta)

static compute_A_nat(eta)

static compute_A_std(mu, Lambda)

static compute_suff_stats(x)

plot1D(feat_idx=0, num_sigmas=2, num_pts=100, **kwargs)

plot2D(feat_idx=[0, 1], num_sigmas=2, num_pts=100, **kwargs)

plot3D(feat_idx=[0, 1], num_sigmas=2, num_pts=100, **kwargs)

plot3D_ellipsoid(feat_idx=[0, 1, 2], num_sigmas=2, num_pts=100, **kwargs)

Estep(x, u_x=None, sample_weight=None, batch_size=None)

Estep_generator(generator, num_steps, return_log_h, max_queue_size=10, workers=1, use_multiprocessin=False)

accum_log_h(x, sample_weight=None)

accum_suff_stats(x, u_x=None, sample_weight=None, batch_size=None)

accum_suff_stats_segments(x, segments, u_x=None, sample_weight=None, batch_size=None)

accum_suff_stats_segments_prob(x, prob, u_x=None, sample_weight=None, batch_size=None)

accum_suff_stats_sorttime(x, frame_length, frame_shift, u_x=None, sample_weight=None, batch_size=None)

compute_log_pz(x, u_x=None, mode='nat')

compute_pz(x, u_x=None, mode='nat')

compute_pz_nat(x, u_x=None)

compute_pz_std(x)

copy()

elbo(x, u_x=None, N=1, log_h=None, sample_weight=None, batch_size=None)

eval_llk(x)

fit(x, sample_weight=None, x_val=None, sample_weight_val=None, epochs=10, batch_size=None)

fit_generator(generator, train_steps, epochs=10, val_data=None, val_steps=0, max_queue_size=10, workers=1, use_multiprocessing=False)

generate(num_samples, **kwargs)

init_to_false()

property is_init

classmethod load(file_path)

classmethod load_config(file_path)

static load_config_from_json(json_str)

log_h(x)

property log_pi

log_prob(x, u_x=None, mode='nat')

log_prob_nat(x, u_x=None)

log_prob_nbest(x, u_x=None, mode='nat', nbest_mode='master', nbest=1)

log_prob_nbest_nat(x, u_x=None, nbest_mode='master', nbest=1)

abstract log_prob_nbest_std(x, nbest_mode='master', nbest=1)

abstract save(file_path)

sum_suff_stats(N, u_x)

to_json(**kwargs)

static tuple2data(data)

class hyperion.pdfs.mixtures.gmm_tied_diag_cov.GMMTiedDiagCov(mu=None, Lambda=None, var_floor=0.001, update_mu=True, update_Lambda=True, **kwargs)

__init__(mu=None, Lambda=None, var_floor=0.001, update_mu=True, update_Lambda=True, **kwargs)

Mstep(N, u_x)

split_comp(K=2)

log_prob_std(x)

log_cdf(x)

sample(num_samples, rng=None, seed=1024)

static compute_eta(mu, Lambda)

static compute_std(eta)

plot1D(feat_idx=0, num_sigmas=2, num_pts=100, **kwargs)

plot2D(feat_idx=[0, 1], num_sigmas=2, num_pts=100, **kwargs)

plot3D(feat_idx=[0, 1], num_sigmas=2, num_pts=100, **kwargs)

plot3D_ellipsoid(feat_idx=[0, 1, 2], num_sigmas=2, num_pts=100, **kwargs)

Estep(x, u_x=None, sample_weight=None, batch_size=None)

Estep_generator(generator, num_steps, return_log_h, max_queue_size=10, workers=1, use_multiprocessin=False)

property Sigma

accum_log_h(x, sample_weight=None)

accum_suff_stats(x, u_x=None, sample_weight=None, batch_size=None)

accum_suff_stats_segments(x, segments, u_x=None, sample_weight=None, batch_size=None)

accum_suff_stats_segments_prob(x, prob, u_x=None, sample_weight=None, batch_size=None)

accum_suff_stats_sorttime(x, frame_length, frame_shift, u_x=None, sample_weight=None, batch_size=None)

property cholLambda

static compute_A_nat(eta)

static compute_A_std(mu, Lambda)

compute_log_pz(x, u_x=None, mode='nat')

compute_pz(x, u_x=None, mode='nat')

compute_pz_nat(x, u_x=None)

compute_pz_std(x)

static compute_suff_stats(x)

copy()

elbo(x, u_x=None, N=1, log_h=None, sample_weight=None, batch_size=None)

eval_llk(x)

fit(x, sample_weight=None, x_val=None, sample_weight_val=None, epochs=10, batch_size=None)

fit_generator(generator, train_steps, epochs=10, val_data=None, val_steps=0, max_queue_size=10, workers=1, use_multiprocessing=False)

generate(num_samples, **kwargs)

get_config()

init_to_false()

initialize(x=None)

property is_init

classmethod load(file_path)

classmethod load_config(file_path)

static load_config_from_json(json_str)

classmethod load_from_kaldi(file_path)

classmethod load_params(f, config)

property logLambda

log_h(x)

property log_pi

log_prob(x, u_x=None, mode='nat')

log_prob_nat(x, u_x=None)

log_prob_nbest(x, u_x=None, mode='nat', nbest_mode='master', nbest=1)

log_prob_nbest_nat(x, u_x=None, nbest_mode='master', nbest=1)

abstract log_prob_nbest_std(x, nbest_mode='master', nbest=1)

norm_suff_stats(N, u_x, return_order2=False)

abstract save(file_path)

save_params(f)

stack_suff_stats(F, S=None)

sum_suff_stats(N, u_x)

to_json(**kwargs)

static tuple2data(data)

unstack_suff_stats(stats)

validate()

Classifiers and Calibrators

Gaussian Classifiers

class hyperion.classifiers.linear_gbe.LinearGBE(mu=None, W=None, update_mu=True, update_W=True, x_dim=1, num_classes=None, balance_class_weight=True, beta=None, nu=None, prior=None, prior_beta=None, prior_nu=None, post_beta=None, post_nu=None, **kwargs)

__init__(mu=None, W=None, update_mu=True, update_W=True, x_dim=1, num_classes=None, balance_class_weight=True, beta=None, nu=None, prior=None, prior_beta=None, prior_nu=None, post_beta=None, post_nu=None, **kwargs)

get_config()

eval_linear(x)

eval_llk(x)

eval_predictive(x)

predict(x, eval_method='linear', normalize=False)

fit(x, class_ids=None, p_theta=None, sample_weight=None)

save_params(f)

classmethod load_params(f, config)

static filter_args(**kwargs)

static filter_train_args(**kwargs)

static add_class_args(parser, prefix=None)

static filter_eval_args(prefix, **kwargs)

static add_eval_args(parser, prefix=None)

static add_argparse_args(parser, prefix=None)

static add_argparse_train_args(parser, prefix=None)

static add_argparse_eval_args(parser, prefix=None)

copy()

abstract fit_generator(x, x_val=None)

init_to_false()

abstract initialize()

property is_init

classmethod load(file_path)

classmethod load_config(file_path)

static load_config_from_json(json_str)

abstract save(file_path)

to_json(**kwargs)

SVM Classifiers

class hyperion.classifiers.linear_svmc.LinearSVMC(A=None, b=None, penalty='l2', C=1.0, loss='squared_hinge', use_bias=True, bias_scaling=1, class_weight=None, random_state=None, max_iter=100, dual=True, tol=0.0001, multi_class='ovr', verbose=0, balance_class_weight=True, lr_seed=1024, **kwargs)

__init__(A=None, b=None, penalty='l2', C=1.0, loss='squared_hinge', use_bias=True, bias_scaling=1, class_weight=None, random_state=None, max_iter=100, dual=True, tol=0.0001, multi_class='ovr', verbose=0, balance_class_weight=True, lr_seed=1024, **kwargs)

property A

property b

get_config()

predict(x, eval_type='logit')

fit(x, class_ids, sample_weight=None)

save_params(f)

classmethod load_params(f, config)

static filter_train_args(prefix=None, **kwargs)

static add_class_train_args(parser, prefix=None)

static filter_eval_args(prefix, **kwargs)

static add_class_eval_args(parser, prefix=None)

static add_argparse_train_args(parser, prefix=None)

static add_argparse_eval_args(parser, prefix=None)

copy()

abstract fit_generator(x, x_val=None)

init_to_false()

abstract initialize()

property is_init

classmethod load(file_path)

classmethod load_config(file_path)

static load_config_from_json(json_str)

abstract save(file_path)

to_json(**kwargs)

Logistic Regression Classifiers and Calibrators

class hyperion.classifiers.logistic_regression.LogisticRegression(A=None, b=None, penalty='l2', lambda_reg=1e-05, use_bias=True, bias_scaling=1, priors=None, random_state=None, solver='lbfgs', max_iter=100, dual=False, tol=0.0001, multi_class='multinomial', verbose=0, warm_start=True, num_jobs=1, lr_seed=1024, **kwargs)

__init__(A=None, b=None, penalty='l2', lambda_reg=1e-05, use_bias=True, bias_scaling=1, priors=None, random_state=None, solver='lbfgs', max_iter=100, dual=False, tol=0.0001, multi_class='multinomial', verbose=0, warm_start=True, num_jobs=1, lr_seed=1024, **kwargs)

Wrapper for sktlearn logistic regression. penalty : str, ‘l1’ or ‘l2’, default: ‘l2’ ,

Used to specify the norm used in the penalization. The ‘newton-cg’, ‘sag’ and ‘lbfgs’ solvers support only l2 penalties.

New in version 0.19: l1 penalty with SAGA solver (allowing ‘multinomial’ + L1)

dualbool, default: False

Dual or primal formulation. Dual formulation is only implemented for l2 penalty with liblinear solver. Prefer dual=False when n_samples > n_features.

tolfloat, default: 1e-4

Tolerance for stopping criteria.

lambda_regfloat, default: 1e-5

Regularization strength; must be a positive float.

use_biasbool, default: True

Specifies if a constant (a.k.a. bias or intercept) should be added to the decision function.

bias_scalingfloat, default 1.

Useful only when the solver ‘liblinear’ is used and use_bias is set to True. In this case, x becomes [x, bias_scaling], i.e. a “synthetic” feature with constant value equal to intercept_scaling is appended to the instance vector. The intercept becomes intercept_scaling * synthetic_feature_weight. Note! the synthetic feature weight is subject to l1/l2 regularization as all other features. To lessen the effect of regularization on synthetic feature weight (and therefore on the intercept) bias_scaling has to be increased.

priorsdict or ‘balanced’ default: None

Weights associated with classes in the form {class_label: weight}. If not given, all classes are supposed to have weight one. The “balanced” mode uses the values of y to automatically adjust weights inversely proportional to class frequencies in the input data as n_samples / (n_classes * np.bincount(y)). Note that these weights will be multiplied with sample_weight (passed through the fit method) if sample_weight is specified.

random_stateint, RandomState instance or None, optional, default: None

The seed of the pseudo random number generator to use when shuffling the data. If int, random_state is the seed used by the random number generator; If RandomState instance, random_state is the random number generator; . Used when solver == ‘sag’ or ‘liblinear’.

solver{‘newton-cg’, ‘lbfgs’, ‘liblinear’, ‘sag’, ‘saga’},

default: ‘liblinear’ Algorithm to use in the optimization problem. For small datasets, ‘liblinear’ is a good choice, whereas ‘sag’ and ‘saga’ are faster for large ones. For multiclass problems, only ‘newton-cg’, ‘sag’, ‘saga’ and ‘lbfgs’ handle multinomial loss; ‘liblinear’ is limited to one-versus-rest schemes. ‘newton-cg’, ‘lbfgs’ and ‘sag’ only handle L2 penalty, whereas ‘liblinear’ and ‘saga’ handle L1 penalty. Note that ‘sag’ and ‘saga’ fast convergence is only guaranteed on features with approximately the same scale. New in version 0.17: Stochastic Average Gradient descent solver. New in version 0.19: SAGA solver.

max_iterint, default: 100

Useful only for the newton-cg, sag and lbfgs solvers. Maximum number of iterations taken for the solvers to converge.

multi_classstr, {‘ovr’, ‘multinomial’}, default: ‘ovr’

Multiclass option can be either ‘ovr’ or ‘multinomial’. If the option chosen is ‘ovr’, then a binary problem is fit for each label. Else the loss minimised is the multinomial loss fit across the entire probability distribution. Does not work for liblinear solver. New in version 0.18: Stochastic Average Gradient descent solver for ‘multinomial’ case.

verboseint, default: 0

For the liblinear and lbfgs solvers set verbose to any positive number for verbosity.

warm_startbool, default: False

When set to True, reuse the solution of the previous call to fit as initialization, otherwise, just erase the previous solution. Useless for liblinear solver. New in version 0.17: warm_start to support lbfgs, newton-cg, sag, saga solvers.

n_jobsint, default: 1

Number of CPU cores used when parallelizing over classes if multi_class=’ovr’”. This parameter is ignored when the ``solver``is set to ‘liblinear’ regardless of whether ‘multi_class’ is specified or not. If given a value of -1, all cores are used.

property A

property b

get_config()

predict(x, eval_type='logit')

fit(x, class_ids, sample_weight=None)

save_params(f)

classmethod load_params(f, config)

static filter_args(prefix=None, **kwargs)

static filter_train_args(prefix=None, **kwargs)

static add_class_args(parser, prefix=None)

static filter_eval_args(prefix, **kwargs)

static add_eval_args(parser, prefix=None)

static add_argparse_args(parser, prefix=None)

static add_argparse_train_args(parser, prefix=None)

static add_argparse_eval_args(parser, prefix=None)

copy()

abstract fit_generator(x, x_val=None)

init_to_false()

abstract initialize()

property is_init

classmethod load(file_path)

classmethod load_config(file_path)

static load_config_from_json(json_str)

abstract save(file_path)

to_json(**kwargs)

class hyperion.classifiers.binary_logistic_regression.BinaryLogisticRegression(A=None, b=None, penalty='l2', lambda_reg=1e-06, use_bias=True, bias_scaling=1, prior=0.5, random_state=None, solver='liblinear', max_iter=100, dual=False, tol=0.0001, verbose=0, warm_start=True, lr_seed=1024, **kwargs)

__init__(A=None, b=None, penalty='l2', lambda_reg=1e-06, use_bias=True, bias_scaling=1, prior=0.5, random_state=None, solver='liblinear', max_iter=100, dual=False, tol=0.0001, verbose=0, warm_start=True, lr_seed=1024, **kwargs)

Wrapper for sktlearn logistic regression. penalty : str, ‘l1’ or ‘l2’, default: ‘l2’ ,

Used to specify the norm used in the penalization. The ‘newton-cg’, ‘sag’ and ‘lbfgs’ solvers support only l2 penalties.

New in version 0.19: l1 penalty with SAGA solver (allowing ‘multinomial’ + L1)

dualbool, default: False

Dual or primal formulation. Dual formulation is only implemented for l2 penalty with liblinear solver. Prefer dual=False when n_samples > n_features.

tolfloat, default: 1e-4

Tolerance for stopping criteria.

lambda_regfloat, default: 1e-5

Regularization strength; must be a positive float.

use_biasbool, default: True

Specifies if a constant (a.k.a. bias or intercept) should be added to the decision function.

bias_scalingfloat, default 1.

Useful only when the solver ‘liblinear’ is used and use_bias is set to True. In this case, x becomes [x, bias_scaling], i.e. a “synthetic” feature with constant value equal to intercept_scaling is appended to the instance vector. The intercept becomes intercept_scaling * synthetic_feature_weight. Note! the synthetic feature weight is subject to l1/l2 regularization as all other features. To lessen the effect of regularization on synthetic feature weight (and therefore on the intercept) bias_scaling has to be increased.

priorsdict or ‘balanced’ default: None

Weights associated with classes in the form {class_label: weight}. If not given, all classes are supposed to have weight one. The “balanced” mode uses the values of y to automatically adjust weights inversely proportional to class frequencies in the input data as n_samples / (n_classes * np.bincount(y)). Note that these weights will be multiplied with sample_weight (passed through the fit method) if sample_weight is specified.

random_stateint, RandomState instance or None, optional, default: None

The seed of the pseudo random number generator to use when shuffling the data. If int, random_state is the seed used by the random number generator; If RandomState instance, random_state is the random number generator; . Used when solver == ‘sag’ or ‘liblinear’.

solver{‘newton-cg’, ‘lbfgs’, ‘liblinear’, ‘sag’, ‘saga’},

default: ‘liblinear’ Algorithm to use in the optimization problem. For small datasets, ‘liblinear’ is a good choice, whereas ‘sag’ and ‘saga’ are faster for large ones. For multiclass problems, only ‘newton-cg’, ‘sag’, ‘saga’ and ‘lbfgs’ handle multinomial loss; ‘liblinear’ is limited to one-versus-rest schemes. ‘newton-cg’, ‘lbfgs’ and ‘sag’ only handle L2 penalty, whereas ‘liblinear’ and ‘saga’ handle L1 penalty. Note that ‘sag’ and ‘saga’ fast convergence is only guaranteed on features with approximately the same scale. New in version 0.17: Stochastic Average Gradient descent solver. New in version 0.19: SAGA solver.

max_iterint, default: 100

Useful only for the newton-cg, sag and lbfgs solvers. Maximum number of iterations taken for the solvers to converge.

multi_classstr, {‘ovr’, ‘multinomial’}, default: ‘ovr’

Multiclass option can be either ‘ovr’ or ‘multinomial’. If the option chosen is ‘ovr’, then a binary problem is fit for each label. Else the loss minimised is the multinomial loss fit across the entire probability distribution. Does not work for liblinear solver. New in version 0.18: Stochastic Average Gradient descent solver for ‘multinomial’ case.

verboseint, default: 0

For the liblinear and lbfgs solvers set verbose to any positive number for verbosity.

warm_startbool, default: False

When set to True, reuse the solution of the previous call to fit as initialization, otherwise, just erase the previous solution. Useless for liblinear solver. New in version 0.17: warm_start to support lbfgs, newton-cg, sag, saga solvers.

n_jobsint, default: 1

Number of CPU cores used when parallelizing over classes if multi_class=’ovr’”. This parameter is ignored when the ``solver``is set to ‘liblinear’ regardless of whether ‘multi_class’ is specified or not. If given a value of -1, all cores are used.

property prior

get_config()

predict(x, eval_type='logit')

static filter_train_args(**kwargs)

static add_class_args(parser, prefix=None)

static add_argparse_args(parser, prefix=None)

property A

static add_argparse_eval_args(parser, prefix=None)

static add_argparse_train_args(parser, prefix=None)

static add_eval_args(parser, prefix=None)

property b

copy()

static filter_args(prefix=None, **kwargs)

static filter_eval_args(prefix, **kwargs)

fit(x, class_ids, sample_weight=None)

abstract fit_generator(x, x_val=None)

init_to_false()

abstract initialize()

property is_init

classmethod load(file_path)

classmethod load_config(file_path)

static load_config_from_json(json_str)

classmethod load_params(f, config)

abstract save(file_path)

save_params(f)

to_json(**kwargs)

Clustering Tools

class hyperion.clustering.kmeans.KMeans(num_clusters, mu=None, rtol=0.001, **kwargs)

__init__(num_clusters, mu=None, rtol=0.001, **kwargs)

fit(x, epochs=100)

predict(x)

copy()

abstract fit_generator(x, x_val=None)

abstract get_config()

init_to_false()

abstract initialize()

property is_init

classmethod load(file_path)

classmethod load_config(file_path)

static load_config_from_json(json_str)

classmethod load_params(f, config)

abstract save(file_path)

abstract save_params(f)

to_json(**kwargs)

class hyperion.clustering.ahc.AHC(method='average', metric='llr', **kwargs)

__init__(method='average', metric='llr', **kwargs)

fit(x, mask=None)

get_flat_clusters(t, criterion='threshold')

get_flat_clusters_from_num_clusters(num_clusters)

get_flat_clusters_from_thr(thr)

compute_flat_clusters()

evaluate_impurity_det(labels_true)

copy()

abstract fit_generator(x, x_val=None)

abstract get_config()

init_to_false()

abstract initialize()

property is_init

classmethod load(file_path)

classmethod load_config(file_path)

static load_config_from_json(json_str)

classmethod load_params(f, config)

abstract save(file_path)

abstract save_params(f)

to_json(**kwargs)

Score Normalization

class hyperion.score_norm.score_norm.ScoreNorm(std_floor=1e-05, **kwargs)

Base class for score normalization

__init__(std_floor=1e-05, **kwargs)

copy()

abstract fit(x, sample_weights=None, x_val=None, sample_weights_val=None)

abstract fit_generator(x, x_val=None)

abstract get_config()

init_to_false()

abstract initialize()

property is_init

classmethod load(file_path)

classmethod load_config(file_path)

static load_config_from_json(json_str)

classmethod load_params(f, config)

abstract save(file_path)

abstract save_params(f)

to_json(**kwargs)

class hyperion.score_norm.t_norm.TNorm(std_floor=1e-05, **kwargs)

Class for T-Norm score normalization.

predict(scores, scores_coh_test, mask=None)

__init__(std_floor=1e-05, **kwargs)

copy()

abstract fit(x, sample_weights=None, x_val=None, sample_weights_val=None)

abstract fit_generator(x, x_val=None)

abstract get_config()

init_to_false()

abstract initialize()

property is_init

classmethod load(file_path)

classmethod load_config(file_path)

static load_config_from_json(json_str)

classmethod load_params(f, config)

abstract save(file_path)

abstract save_params(f)

to_json(**kwargs)

class hyperion.score_norm.z_norm.ZNorm(std_floor=1e-05, **kwargs)

Class for Z-Norm score normalization.

predict(scores, scores_enr_coh, mask=None)

__init__(std_floor=1e-05, **kwargs)

copy()

abstract fit(x, sample_weights=None, x_val=None, sample_weights_val=None)

abstract fit_generator(x, x_val=None)

abstract get_config()

init_to_false()

abstract initialize()

property is_init

classmethod load(file_path)

classmethod load_config(file_path)

static load_config_from_json(json_str)

classmethod load_params(f, config)

abstract save(file_path)

abstract save_params(f)

to_json(**kwargs)

class hyperion.score_norm.zt_norm.ZTNorm(**kwargs)

Class ZT-Norm score-normalization.

__init__(**kwargs)

predict(scores, scores_coh_test, scores_enr_coh, scores_coh_coh, mask_coh_test=None, mask_enr_coh=None, mask_coh_coh=None)

copy()

abstract fit(x, sample_weights=None, x_val=None, sample_weights_val=None)

abstract fit_generator(x, x_val=None)

abstract get_config()

init_to_false()

abstract initialize()

property is_init

classmethod load(file_path)

classmethod load_config(file_path)

static load_config_from_json(json_str)

classmethod load_params(f, config)

abstract save(file_path)

abstract save_params(f)

to_json(**kwargs)

class hyperion.score_norm.tz_norm.TZNorm(**kwargs)

Class for TZ-Norm score normalization.

__init__(**kwargs)

predict(scores, scores_coh_test, scores_enr_coh, scores_coh_coh, mask_coh_test=None, mask_enr_coh=None, mask_coh_coh=None)

copy()

abstract fit(x, sample_weights=None, x_val=None, sample_weights_val=None)

abstract fit_generator(x, x_val=None)

abstract get_config()

init_to_false()

abstract initialize()

property is_init

classmethod load(file_path)

classmethod load_config(file_path)

static load_config_from_json(json_str)

classmethod load_params(f, config)

abstract save(file_path)

abstract save_params(f)

to_json(**kwargs)

class hyperion.score_norm.s_norm.SNorm(**kwargs)

Class for S-Norm, symmetric score normalization.

__init__(**kwargs)

predict(scores, scores_coh_test, scores_enr_coh, mask_coh_test=None, mask_enr_coh=None)

copy()

abstract fit(x, sample_weights=None, x_val=None, sample_weights_val=None)

abstract fit_generator(x, x_val=None)

abstract get_config()

init_to_false()

abstract initialize()

property is_init

classmethod load(file_path)

classmethod load_config(file_path)

static load_config_from_json(json_str)

classmethod load_params(f, config)

abstract save(file_path)

abstract save_params(f)

to_json(**kwargs)

class hyperion.score_norm.adapt_s_norm.AdaptSNorm(nbest=100, nbest_discard=0, **kwargs)

Class for adaptive S-Norm

__init__(nbest=100, nbest_discard=0, **kwargs)

predict(scores, scores_coh_test, scores_enr_coh, mask_coh_test=None, mask_enr_coh=None)

copy()

abstract fit(x, sample_weights=None, x_val=None, sample_weights_val=None)

abstract fit_generator(x, x_val=None)

abstract get_config()

init_to_false()

abstract initialize()

property is_init

classmethod load(file_path)

classmethod load_config(file_path)

static load_config_from_json(json_str)

classmethod load_params(f, config)

abstract save(file_path)

abstract save_params(f)

to_json(**kwargs)

Feature Transformations

These are classes to apply feature transformations/projections like PCA, LDA, etc.

Transform Classes

class hyperion.transforms.pca.PCA(mu=None, T=None, update_mu=True, update_T=True, pca_dim=None, pca_var_r=None, pca_min_dim=2, whiten=False, **kwargs)

Class to do principal component analysis

__init__(mu=None, T=None, update_mu=True, update_T=True, pca_dim=None, pca_var_r=None, pca_min_dim=2, whiten=False, **kwargs)

predict(x)

static get_pca_dim_for_var_ratio(x, var_r=1, min_dim=2)

fit(x=None, sample_weight=None, mu=None, S=None)

get_config()

save_params(f)

classmethod load_params(f, config)

classmethod load_mat(file_path)

save_mat(file_path)

static filter_args(**kwargs)

static add_class_args(parser, prefix=None)

static add_argparse_args(parser, prefix=None)

copy()

abstract fit_generator(x, x_val=None)

init_to_false()

abstract initialize()

property is_init

classmethod load(file_path)

classmethod load_config(file_path)

static load_config_from_json(json_str)

abstract save(file_path)

to_json(**kwargs)

class hyperion.transforms.lda.LDA(mu=None, T=None, lda_dim=None, update_mu=True, update_T=True, **kwargs)

Class to do linear discriminant analysis.

__init__(mu=None, T=None, lda_dim=None, update_mu=True, update_T=True, **kwargs)

predict(x)

fit(x, y, mu=None, Sb=None, Sw=None)

get_config()

save_params(f)

classmethod load_params(f, config)

classmethod load_mat(file_path)

copy()

abstract fit_generator(x, x_val=None)

init_to_false()

abstract initialize()

property is_init

classmethod load(file_path)

classmethod load_config(file_path)

static load_config_from_json(json_str)

abstract save(file_path)

save_mat(file_path)

to_json(**kwargs)

class hyperion.transforms.cent_whiten.CentWhiten(mu=None, T=None, update_mu=True, update_T=True, **kwargs)

Class to do centering and whitening of i-vectors.

__init__(mu=None, T=None, update_mu=True, update_T=True, **kwargs)

predict(x)

fit(x=None, sample_weight=None, mu=None, S=None)

get_config()

save_params(f)

classmethod load_params(f, config)

classmethod load_mat(file_path)

save_mat(file_path)

static filter_args(**kwargs)

static add_class_args(parser, prefix=None)

static add_argparse_args(parser, prefix=None)

copy()

abstract fit_generator(x, x_val=None)

init_to_false()

abstract initialize()

property is_init

classmethod load(file_path)

classmethod load_config(file_path)

static load_config_from_json(json_str)

abstract save(file_path)

to_json(**kwargs)

class hyperion.transforms.lnorm.LNorm(mu=None, T=None, update_mu=True, update_T=True, **kwargs)

Class to do length normalization.

predict(x)

__init__(mu=None, T=None, update_mu=True, update_T=True, **kwargs)

static add_argparse_args(parser, prefix=None)

static add_class_args(parser, prefix=None)

copy()

static filter_args(**kwargs)

fit(x=None, sample_weight=None, mu=None, S=None)

abstract fit_generator(x, x_val=None)

get_config()

init_to_false()

abstract initialize()

property is_init

classmethod load(file_path)

classmethod load_config(file_path)

static load_config_from_json(json_str)

classmethod load_mat(file_path)

classmethod load_params(f, config)

abstract save(file_path)

save_mat(file_path)

save_params(f)

to_json(**kwargs)

class hyperion.transforms.coral.CORAL(mu=None, T_col=None, T_white=None, update_mu=True, update_T=True, alpha_mu=1, alpha_T=1, **kwargs)

Class to do CORAL

__init__(mu=None, T_col=None, T_white=None, update_mu=True, update_T=True, alpha_mu=1, alpha_T=1, **kwargs)

get_config()

predict(x)

fit(x, sample_weight=None, x_out=None, sample_weight_out=None)

classmethod load_params(f, config)

save_params(f)

copy()

abstract fit_generator(x, x_val=None)

init_to_false()

abstract initialize()

property is_init

classmethod load(file_path)

classmethod load_config(file_path)

static load_config_from_json(json_str)

abstract save(file_path)

to_json(**kwargs)

class hyperion.transforms.gaussianizer.Gaussianizer(max_vectors=None, r=None, **kwargs)

Class to make i-vector distribution standard Normal.

__init__(max_vectors=None, r=None, **kwargs)

predict(x)

fit(x)

get_config()

save_params(f)

classmethod load_params(f, config)

classmethod load_mat(file_path)

save_mat(file_path)

static filter_args(**kwargs)

static add_class_args(parser, prefix=None)

static add_arparse_args(parser, prefix=None)

copy()

abstract fit_generator(x, x_val=None)

init_to_false()

abstract initialize()

property is_init

classmethod load(file_path)

classmethod load_config(file_path)

static load_config_from_json(json_str)

abstract save(file_path)

to_json(**kwargs)

class hyperion.transforms.nap.NAP(U=None, **kwargs)

Class to do nussance attribute projection.

__init__(U=None, **kwargs)

predict(x)

fit(x, U_dim, class_ids)

save_params(f)

classmethod load_params(f, config)

classmethod load_mat(file_path)

save_mat(file_path)

copy()

abstract fit_generator(x, x_val=None)

abstract get_config()

init_to_false()

abstract initialize()

property is_init

classmethod load(file_path)

classmethod load_config(file_path)

static load_config_from_json(json_str)

abstract save(file_path)

to_json(**kwargs)

class hyperion.transforms.nda.NDA(mu=None, T=None, **kwargs)

Class to do nearest-neighbors discriminant analysis

__init__(mu=None, T=None, **kwargs)

predict(x)

fit(mu, Sb, Sw, nda_dim=None)

save_params(f)

classmethod load_params(f, config)

classmethod load_mat(file_path)

save_mat(file_path)

copy()

abstract fit_generator(x, x_val=None)

abstract get_config()

init_to_false()

abstract initialize()

property is_init

classmethod load(file_path)

classmethod load_config(file_path)

static load_config_from_json(json_str)

abstract save(file_path)

to_json(**kwargs)

class hyperion.transforms.mvn.MVN(mu=None, s=None, **kwargs)

Class to do global mean and variance normalization.

__init__(mu=None, s=None, **kwargs)

predict(x)

fit(x)

save_params(f)

classmethod load_params(f, config)

copy()

abstract fit_generator(x, x_val=None)

abstract get_config()

init_to_false()

abstract initialize()

property is_init

classmethod load(file_path)

classmethod load_config(file_path)

static load_config_from_json(json_str)

abstract save(file_path)

to_json(**kwargs)

class hyperion.transforms.skl_tsne.SklTSNE(tsne_dim=2, perplexity=30.0, early_exaggeration=12.0, lr=200.0, num_iter=1000, num_iter_without_progress=300, min_grad_norm=1e-07, metric='euclidean', init='random', verbose=0, rng=None, rng_seed=1234, method='barnes_hut', angle=0.5, num_jobs=None, **kwargs)

Wrapper class for sklearn TSNE manifold learner

tsne_dim

dimension of the embedded space.

perplexity

the perplexity is related to the number of nearest neighbors that is used in other manifold learning algorithms. Larger datasets usually require a larger perplexity. Consider selecting a value between 5 and 50.

early_exaggeration

controls how tight natural clusters in the original space are in the embedded space and how much space will be between them.

lr

the learning rate for t-SNE is usually in the range [10.0, 1000.0].

num_iter

maximum number of iterations for the optimization.

num_iter_without_progress

maximum number of iterations without progress before we abort the optimization

min_grad_norm

if the gradient norm is below this threshold, the optimization will be stopped.

metric

the metric to use when calculating distance between instances in [‘cosine’, ‘euclidean’, ‘l1’, ‘l2’, ‘precomputed’] or callable function.

init

initialization method in [‘random’, ‘pca’] or embedding matrix of shape (num_samples, num_comp)

verbose

verbosity level.

rng

RandomState instance

rng_seed

seed for random number generator

method

gradient calculation method in [‘barnes_hut’, ‘exact’]

angle

angle thetha in Barnes-Hut TSNE

num_jobs

number of parallel jobs to run for neighbors search.

__init__(tsne_dim=2, perplexity=30.0, early_exaggeration=12.0, lr=200.0, num_iter=1000, num_iter_without_progress=300, min_grad_norm=1e-07, metric='euclidean', init='random', verbose=0, rng=None, rng_seed=1234, method='barnes_hut', angle=0.5, num_jobs=None, **kwargs)

property tsne_dim

property perplexity

property early_exaggeration

property lr

property num_iter

property num_iter_without_progress

property min_grad_norm

property metric

property init

property method

property angle

property num_jobs

predict(x)

fit(x)

save_params(f)

classmethod load_params(f, config)

get_config()

static filter_args(**kwargs)

static add_class_args(parser, prefix=None)

static add_argparse_args(parser, prefix=None)

copy()

abstract fit_generator(x, x_val=None)

init_to_false()

abstract initialize()

property is_init

classmethod load(file_path)

classmethod load_config(file_path)

static load_config_from_json(json_str)

abstract save(file_path)

to_json(**kwargs)

Sequence of Transformations Classes

class hyperion.transforms.transform_list.TransformList(transforms, **kwargs)

Class to perform a list of transformations

__init__(transforms, **kwargs)

append(t)

predict(x)

update_names()

get_config()

save_params(f)

classmethod load_params(f, config)

copy()

abstract fit(x, sample_weights=None, x_val=None, sample_weights_val=None)

abstract fit_generator(x, x_val=None)

init_to_false()

abstract initialize()

property is_init

classmethod load(file_path)

classmethod load_config(file_path)

static load_config_from_json(json_str)

abstract save(file_path)

to_json(**kwargs)

Auxiliary Classes

class hyperion.transforms.sb_sw.SbSw(Sb=None, Sw=None, mu=None, num_classes=0, **kwargs)

Class to compute between and within class matrices

__init__(Sb=None, Sw=None, mu=None, num_classes=0, **kwargs)

fit(x, class_ids, sample_weight=None, class_weights=None, normalize=True)

normalize()

classmethod accum_stats(stats)

save_params(f)

classmethod load(file_path)

copy()

abstract fit_generator(x, x_val=None)

abstract get_config()

init_to_false()

abstract initialize()

property is_init

classmethod load_config(file_path)

static load_config_from_json(json_str)

classmethod load_params(f, config)

abstract save(file_path)

to_json(**kwargs)

Metrics

Metric Functions

These are some functions to compute performance metrics used in speaker identification and verification

Copyright 2018 Johns Hopkins University (Author: Jesus Villalba) Apache 2.0 (http://www.apache.org/licenses/LICENSE-2.0)

hyperion.metrics.eer.compute_eer(tar, non)

Computes equal error rate.

Parameters

• tar – Scores of target trials.

• non – Scores of non-target trials.

Returns

EER

hyperion.metrics.eer.compute_prbep(tar, non)

Computes precission-recall break-even point

where #FA == #Miss

Parameters

• tar – Scores of target trials.

• non – Scores of non-target trials.

Returns

PREBP value

Copyright 2018 Johns Hopkins University (Author: Jesus Villalba) Apache 2.0 (http://www.apache.org/licenses/LICENSE-2.0)

hyperion.metrics.dcf.compute_dcf(p_miss, p_fa, prior, normalize=True)

Computes detection cost function

DCF = prior*p_miss + (1-prior)*p_fa

Parameters

• p_miss – Vector of miss probabilities.

• p_fa – Vector of false alarm probabilities.

• prior – Target prior or vector of target priors.

• normalize – if true, return normalized DCF, else unnormalized.

Returns

Matrix of DCF for each pair of (p_miss, p_fa) and each value of prior. [len(prior) x len(p_miss)]

hyperion.metrics.dcf.compute_min_dcf(tar, non, prior, normalize=True)

Computes minimum DCF

min_DCF = min_t prior*p_miss(t) + (1-prior)*p_fa(t)

where t is the decission threshold.

Parameters

• tar – Target scores.

• non – Non-target scores.

• prior – Target prior or vector of target priors.

• normalize – if true, return normalized DCF, else unnormalized.

Returns

Vector Minimum DCF for each prior. Vector of P_miss corresponding to each min DCF. Vector of P_fa corresponding to each min DCF.

hyperion.metrics.dcf.compute_act_dcf(tar, non, prior, normalize=True)

Computes actual DCF by making decisions assuming that scores

are calibrated to act as log-likelihood ratios.

Parameters

• tar – Target scores.

• non – Non-target scores.

• prior – Target prior or vector of target priors.

• normalize – if true, return normalized DCF, else unnormalized.

Returns

Vector actual DCF for each prior. Vector of P_miss corresponding to each act DCF. Vector of P_fa corresponding to each act DCF.

hyperion.metrics.dcf.fast_eval_dcf_eer(tar, non, prior, normalize_dcf=True, return_probs=False)

Computes actual DCF, minimum DCF, EER and PRBE all togther

Parameters

• tar – Target scores.

• non – Non-target scores.

• prior – Target prior or vector of target priors.

• normalize_cdf – if true, return normalized DCF, else unnormalized.

Returns

Vector Minimum DCF for each prior. Vector Actual DCF for each prior. EER value PREBP value

Copyright 2018 Johns Hopkins University (Author: Jesus Villalba) Apache 2.0 (http://www.apache.org/licenses/LICENSE-2.0)

hyperion.metrics.cllr.compute_cllr(tar, non)

CLLR: Measure of goodness of log-likelihood-ratio detection output. This measure ps both:

• The quality of the score (over the whole DET curve), and

• The quality of the calibration

Parameters

• tar – Scores of target trials.

• non – Scores of non-target trials.

Returns

CLLR

hyperion.metrics.cllr.compute_min_cllr(tar, non)

Copyright 2018 Johns Hopkins University (Author: Jesus Villalba) Apache 2.0 (http://www.apache.org/licenses/LICENSE-2.0)

hyperion.metrics.roc.compute_roc(true_scores, false_scores)

Computes the (observed) miss/false_alarm probabilities

for a set of detection output scores.

Parameters

• true_scores (false_scores) –

• trials (detection) – (By convention, the more positive the score, the more likely is the target hypothesis.)

• true (given that the target hypothesis is) – (By convention, the more positive the score, the more likely is the target hypothesis.)

Returns

The miss/false_alarm error probabilities

hyperion.metrics.roc.compute_rocch(tar_scores, non_scores)

Computes ROCCH: ROC Convex Hull.

Parameters

• tar_scores – scores for target trials

• nontar_scores – scores for non-target trials

Returns

pmiss and pfa contain the coordinates of the vertices of the ROC Convex Hull.

hyperion.metrics.roc.rocch2eer(p_miss, p_fa)

Calculates the equal error rate (eer) from pmiss and pfa vectors. Note: pmiss and pfa contain the coordinates of the vertices of the ROC Convex Hull. Use compute_rocch to convert target and non-target scores to pmiss and pfa values.

hyperion.metrics.roc.filter_roc(p_miss, p_fa)

Removes redundant points from the sequence of points (p_fa,p_miss) so

that plotting an ROC or DET curve will be faster. The output ROC curve will be identical to the one plotted from the input vectors. All points internal to straight (horizontal or vertical) sections on the ROC curve are removed i.e. only the points at the start and end of line segments in the curve are retained. Since the plotting code draws straight lines between points, the resulting plot will be the same as the original.

Parameters

• p_miss – The coordinates of the vertices of the ROC Convex Hull. m for misses and fa for false alarms.

• p_fa – The coordinates of the vertices of the ROC Convex Hull. m for misses and fa for false alarms.

Returns

Vectors containing selected values from the

input vectors.

Return type

new_p_miss, new_p_fa

hyperion.metrics.roc.compute_area_under_rocch(p_miss, p_fa)

Calculates area under the ROC convex hull given p_miss, p_fa.

Parameters

• p_miss – Miss probabilities vector obtained from compute_rocch

• p_fa – False alarm probabilities vector

Returns

AUC

hyperion.metrics.roc.test_roc()

Copyright 2018 Johns Hopkins University (Author: Jesus Villalba) Apache 2.0 (http://www.apache.org/licenses/LICENSE-2.0)

hyperion.metrics.acc.compute_accuracy(y_true, y_pred, normalize=True, sample_weight=None)

Computes accuracy

Parameters

• y_true – 1d array-like, or label indicator array / sparse matrix. Ground truth (correct) labels.

• y_pred – 1d array-like, or label indicator array / sparse matrix. Predicted labels, as returned by a classifier.

• normalize – If False, return the number of correctly classified samples. Otherwise, return the fraction of correctly classified samples.

• sample_weight – Sample weights.

Returns

Accuracy or number of correctly classified samples.

Copyright 2018 Johns Hopkins University (Author: Jesus Villalba) Apache 2.0 (http://www.apache.org/licenses/LICENSE-2.0)

hyperion.metrics.confusion_matrix.compute_confusion_matrix(y_true, y_pred, labels=None, normalize=True, sample_weight=None)

Computes confusion matrix.

Parameters

• y_true – Ground truth.

• y_pred – Estimated labels.

• labels – List of labels to index the matrix. This may be used to reorder or select a subset of labels. If none is given, those that appear at least once in y_true or y_pred are used in sorted order.

• sample_weight – Sample weights.

Returns

Confusion matrix (num_classes x num_classes)

hyperion.metrics.confusion_matrix.compute_xlabel_confusion_matrix(y_true, y_pred, labels_train=None, labels_test=None, normalize=True, sample_weight=None)

Computes confusion matrix when the labels used to train the classifier are

different than those of the test set.

Parameters

• y_true – Ground truth.

• y_pred – Estimated labels.

• labels_train – List of labels used to train the classifier. This may be used to reorder or select a subset of labels. If none is given, those that appear at least once in y_pred are used in sorted order.

• labels_test – List of labels of the test set. This may be used to reorder or select a subset of labels. If none is given, those that appear at least once in y_true are used in sorted order.

• sample_weight – Sample weights.

Returns

Confusion matrix (num_classes_test x num_classes_train)

hyperion.metrics.confusion_matrix.plot_confusion_matrix(C, labels_true, labels_pred=None, title='Confusion matrix', cmap=<matplotlib.colors.LinearSegmentedColormap object>, fmt=None)

Plots a confusion matrix in a figure.

Parameters

• C – 2D numpy array with confusion matrix.

• labels_true – Labels of the true classes (rows).

• labels_cols – Labels of the predicted classes. If None, it is equal to labels_true.

• title – Title for the figure.

• cmp – Color MAP.

hyperion.metrics.confusion_matrix.write_confusion_matrix(f, C, labels_true, labels_pred=None, fmt=None)

Writes confusion matrix to file.

Parameters

• f – Python file hangle.

• C – 2D numpy array with confusion matrix.

• labels_true – Labels of the true classes (rows).

• labels_cols – Labels of the predicted classes. If None, it is equal to labels_true.

hyperion.metrics.confusion_matrix.print_confusion_matrix(C, labels_true, labels_pred=None, fmt=None)

Prints confusion matrix to std output.

Parameters

• C – 2D numpy array with confusion matrix.

• labels_true – Labels of the true classes (rows).

• labels_cols – Labels of the predicted classes. If None, it is equal to labels_true.

Copyright 2018 Johns Hopkins University (Author: Jesus Villalba) Apache 2.0 (http://www.apache.org/licenses/LICENSE-2.0)

Utility functions to evaluate performance

hyperion.metrics.utils.effective_prior(p_tar, c_miss, c_fa)

This function adjusts a given prior probability of target p_targ, to incorporate the effects of a cost of miss, cmiss, and a cost of false-alarm, cfa.

Parameters

• p_tar – target prior

• c_miss – cost of miss

• c_fa – cost of false alarm

Returns

Effective prior

hyperion.metrics.utils.pavx(y)

PAV: Pool Adjacent Violators algorithm. Non-paramtetric optimization subject to monotonicity.

ghat = pav(y) fits a vector ghat with nondecreasing components to the data vector y such that sum((y - ghat).^2) is minimal. (Pool-adjacent-violators algorithm).

Author: This code is and adaptation from Bosaris Toolkit and

it is a simplified version of the ‘IsoMeans.m’ code made available by Lutz Duembgen at:

http://www.imsv.unibe.ch/~duembgen/software

Parameters

y – uncalibrated scores

Returns

Calibrated scores Width of pav bins, from left to right

(the number of bins is data dependent)

Height: corresponding heights of bins (in increasing order)

hyperion.metrics.utils.opt_loglr(tar, non, method='laplace')

Non-parametric optimization of score to log-likelihood-ratio mapping.

Taken from Bosaris toolkit.

Niko Brummer and Johan du Preez, Application-Independent Evaluation of Speaker Detection, Computer Speech and Language, 2005

Parameters

• tar – target scores.

• non – non-target scores.

• method – laplace(default, avoids inf log-LR)/raw

Returns

Calibrated tar and non-tar log-LR

Helper Code Blocks

Classes and codeblocks that are re-used in several scripts

class hyperion.helpers.vector_reader.VectorReader(v_file, key_file, preproc=None, vlist_sep=' ')

Class to load data to train PCA, centering, whitening.

__init__(v_file, key_file, preproc=None, vlist_sep=' ')

read()

static filter_args(**kwargs)

static add_class_args(parser, prefix=None)

static add_argparse_args(parser, prefix=None)

class hyperion.helpers.vector_class_reader.VectorClassReader(v_file, key_file, preproc=None, vlist_sep=' ', class2int_file=None, min_spc=1, max_spc=None, spc_pruning_mode='random', csplit_min_spc=1, csplit_max_spc=None, csplit_mode='random', csplit_overlap=0, vcr_seed=1024, csplit_once=True)

Class to load data to train LDA, PLDA, PDDA.

__init__(v_file, key_file, preproc=None, vlist_sep=' ', class2int_file=None, min_spc=1, max_spc=None, spc_pruning_mode='random', csplit_min_spc=1, csplit_max_spc=None, csplit_mode='random', csplit_overlap=0, vcr_seed=1024, csplit_once=True)

read(return_3d=False, max_length=0)

property class_names

property samples_per_class

property max_samples_per_class

static filter_args(**kwargs)

static add_class_args(parser, prefix=None)

static add_argparse_args(parser, prefix=None)

class hyperion.helpers.trial_data_reader.TrialDataReader(v_file, ndx_file, enroll_file, test_file=None, preproc=None, model_part_idx=1, num_model_parts=1, seg_part_idx=1, num_seg_parts=1, eval_set='enroll-test', tlist_sep=' ')

Loads Ndx, enroll file and x-vectors to evaluate PLDA.

__init__(v_file, ndx_file, enroll_file, test_file=None, preproc=None, model_part_idx=1, num_model_parts=1, seg_part_idx=1, num_seg_parts=1, eval_set='enroll-test', tlist_sep=' ')

read()

static filter_args(**kwargs)

static add_class_args(parser, prefix=None)

static add_argparse_args(parser, prefix=None)

class hyperion.helpers.multi_test_trial_data_reader.MultiTestTrialDataReader(v_file, ndx_file, enroll_file, test_file, test_subseg2orig_file, preproc, tlist_sep=' ', model_idx=1, num_model_parts=1, seg_idx=1, num_seg_parts=1, eval_set='enroll-test')

Loads Ndx, enroll file and x-vectors to evaluate PLDA.

__init__(v_file, ndx_file, enroll_file, test_file, test_subseg2orig_file, preproc, tlist_sep=' ', model_idx=1, num_model_parts=1, seg_idx=1, num_seg_parts=1, eval_set='enroll-test')

read()

static filter_args(**kwargs)

static add_class_args(parser, prefix=None)

static add_argparse_args(parser, prefix=None)

class hyperion.helpers.multi_test_trial_data_reader_v2.MultiTestTrialDataReaderV2(enroll_v_file, test_v_file, ndx_file, enroll_file, test_file, preproc=None, tlist_sep=' ', model_idx=1, num_model_parts=1, seg_idx=1, num_seg_parts=1, eval_set='enroll-test')

Loads Ndx, enroll file and x-vectors to evaluate PLDA.

__init__(enroll_v_file, test_v_file, ndx_file, enroll_file, test_file, preproc=None, tlist_sep=' ', model_idx=1, num_model_parts=1, seg_idx=1, num_seg_parts=1, eval_set='enroll-test')

read()

static filter_args(prefix=None, **kwargs)

static add_argparse_args(parser, prefix=None)