The DKU System for the Speaker Recognition Task of the 2019 VOiCES from a Distance Challenge

The DKU System f or the Speaker Recognition T ask of the 2019 V OiCES fr om a Distance Challenge Danwei Cai 1 , Xiaoyi Qin 1 , 2 , W eicheng Cai 1 , 2 , Ming Li 1 1 Data Science Research Center , Duke Kunshan Uni versity , K unshan, China 2 School of Electronics and Information T echnology , Sun Y at-sen Uni versity , Guangzhou, China ming.li369@dukekunshan.edu.cn Abstract In this paper , we present the DKU system for the speak er recog- nition task of the VOiCES from a distance challenge 2019. W e in vestigate the whole system pipeline for the far-field speaker verification, including data pre-processing, short-term spectral feature representation, utterance-level speaker modeling, back- end scoring, and score normalization. Our best single system employs a residual neural netw ork trained with angular softmax loss. Also, the weighted prediction error algorithms can further improv e performance. It achieves 0.3668 minDCF and 5.58% EER on the evaluation set by using a simple cosine similarity scoring. Finally , the submitted primary system obtains 0.3532 minDCF and 4.96% EER on the ev aluation set. Index T erms : speaker recognition, far -field speech, deep ResNet, angular softmax, WPE 1. Introduction In the past decade, the performance of speaker recognition has improv ed significantly . The i-vector based method [1] and the deep neural network (DNN) based methods [2, 3] have promoted the de velopment of speaker recognition technology in telephone channel and closed talking scenarios. Howe ver , speaker recognition under far-field and complex environmental settings is still challenging due to the effects of the long-range fading, room re verberation, and comple x en vironmental noises. Speech signal propagating in long-range suffers from fading, absorption, and reflection by various objects, which change the pressure le vel at different frequencies and degrade the signal quality [4]. Re verberation includes eaarlay re verberation and late reverberation. Early reverberation (i.e., reflections within 50 to 100 ms after the direct wave arri ves at the microphone) can improv e the recei ved speech quality , while late reverberation will degrade the speech quality . The adverse effects of rev er- beration on speech signal includes smearing spectro-temporal structures, amplifying the lo w-frequency energy , and flattening the formant transitions, etc. [5]. Also, the complex environ- mental noises “fill in” regions with low speech ener gy in the time-frequency plane and blur the spectral details [4]. These ef- fects result in the loss of speech intelligibility and speech qual- ity , imposing great challenges in f ar-field speaker recognition and far -field speech recognition. T o compensate for the adverse impacts of room rev er- beration and en vironmental noise, various approaches hav e This research was funded in part by the National Natural Sci- ence F oundation of China (61773413), Natural Science Foundation of Guangzhou City (201707010363), Six talent peaks project in Jiangsu Province (JY -074), Science and T echnology Program of Guangzhou City (201903010040), and Huawei. W e also thank W eixiang Hu, Y u Lu, Zexin Liu and Lei Miao from Huawei Digital T echnologies Co., Ltd, China. been proposed at different stages of the speaker recognition system. At the signal lev el, dereverberation [6], denoising [7, 8, 9, 10], and beamforming [11, 12] can be used for speech enhancement. At feature lev el, sub-band Hilbert env elopes based features [13, 14], warped minimum variance distortion- less response (MVDR) cepstral coefficients [15], blind spec- tral weighting (BSW) based features [16] hav e been applied to ASV system to suppress the adverse impacts of re verber - ation and noise. At the model lev el, re verberation matching with multi-condition training models has been successfully em- ployed within the uni versal background model (UBM) or i- vector based front-end systems [17, 18]. In back-end mod- eling, multi-condition training of probabilistic linear discrimi- nant analysis (PLDA) models were employed in i-v ector sys- tem [19]. The robustness of deep speaker embeddings for far- field speech has also been in vestigated in [20]. Finally , at the score level, score normalization [17] and multi-channel score fusion [21, 22] hav e been applied in far-field ASV system to improv e the robustness. The “VOiCES from a Distance Challenge 2019” is de- signed to foster research in the area of speaker recognition and automatic speech recognition (ASR) with the special focus on single channel far-eld audio, under noisy conditions [23]. Our system pipeline consists of the following six main components, including data pre-processing, short-term spectral feature ex- traction, utterance-lev el speaker modeling, back-end scoring, score normalization, as well as fusion and calibration. This paper is organized as follows: Section 2 describes the details of our submitted system. Section 3 clarifies the data us- age, with e xperimental results and analysis. Conclusions are drawn in section 4. 2. System descriptions 2.1. Data pre-processing 2.1.1. Data augmentation W e adopt two kinds of data augmentation strategies. The first is the same as the x-vector system av ailable at Kaldi V oxceleb recipe, which employs additive noises and reverberation. W e also use pyr oomacoustics [24] to simulate the room acoustic based on RIR generator using Image Source Model (ISM) al- gorithm. The microphones, distractors, and speech source are similar to the room settings presented in [25]. W e use the mu- sic and noise part of the MUSAN dataset [26] to generate the television noise, and the ‘us-go v’ part to create babble noise. For the systems described below , we use the Kaldi data augmentation strategy for the MFCC i-vector system and the TDNN x-vector system, and pyr oomacoustics data augmenta- tion strategy for the remaining systems. 2.1.2. Dere verberation The weighted prediction error (WPE) algorithm is a success- ful algorithm to reduce late rev erberation [6]. The method es- timates the optimal dereverberation filter coefficients based on iterativ e optimization. During the enrolling and testing, we use the single-channel WPE to dere verberate the sound with a dere- verberation filter of 10 coefficients. The WPE codes are from http://www.kecl.ntt.co.jp/icl/signal/wpe . 2.2. Short-term spectral feature Four feature s including Mel-frequenc y cepstral coef fi- cient (MFCC), po wer-normalized cepstral coef ficients (PNCC), Mel-filterbank energies (Mfbank) and gammatone-Filterbank energies (Gfbank) are adopted in our systems. 2.2.1. MFCC T wo kinds of MFCC features with a different number of cepstral filterbanks are adopted, which result in 20- and 30-dimensional MFCCs (MFCC-20 and MFCC-30). MFCC-20 is for the i- vector system, and MFCC-30 is for the TDNN x-v ector system. Short-time cepstral mean subtraction (CMS) ov er a 3-second sliding window is applied. For the MFCC-20, their first and second deriv atives are computed before applying the CMS. 2.2.2. PNCC PNCC has proved to be more robust in various types of addi- tiv e noise and reverberant en vironments compared to MFCC in ASR [27]. The major features of PNCC processing include the use of a power -law nonlinearity that replaces the traditional log nonlinearity used in MFCC coefficients, a noise-suppression algorithm based on asymmetric filtering that suppress back- ground excitation, and a module that accomplishes temporal masking [27]. 20-dimensional PNCC are extracted using a 25 ms window with 10 ms shifts. First and second deriv ativ es are computed before applying CMS. 2.2.3. Log Mel-filterbank energies Each audio is con verted to 64-dimensional log Mel-filterbank energies with cepstral filterbanks ranging from 20 to 7600 Hz (Mfbank-16k). W e also downsample the audio to 8000 sample rate and use cepstral filter banks within the range of 20 to 3800 Hz to calculate Mfbank-8k features. A short-time cepstral mean subtraction is applied ov er a 3-second sliding window . 2.2.4. Gammatone-Filterbank Ener gies Gammatone filters are approximations to the filtering system of human ear [28]. The Gammatone filterbanks are selected within the range of 50 to 8000 Hz to compute the 64-dimensional Gammatone-filterbank ener gies. Short-time CMS is then ap- plied ov er a 3-second sliding window . 2.3. Utterance-level speaker modeling W e extract the utterance-level speaker embeddings from three state-of-the-art modelings, including the i-vector system [1], the TDNN x-vector system [2], and the deep ResNet system [3]. 2.3.1. i-vector W e train two i-vector systems on the MFCC-20 and PNCC features respecti vely . The extracted 60-dimensional features are used to train a 2048 component Gaussian mixture model- univ ersal background model (GMM-UBM) with full cov ariance matrices. Then zero-order and first-order Baum-W elch statistics are computed on the UBM for each recording’ s MFCC feature, and single factor analysis is employed to extract i-vectors with 600 dimensions [1]. 2.3.2. TDNN x-vector The x-vector system is developed by adapting the Kaldi V ox- celeb recipe. For the x-vector extractor , a DNN is trained to discriminate speakers in the training set. The first fi ve timed delayed layers operate at frame-lev el. Then a temporal statis- tics pooling layer is employed to compute the mean and stan- dard deviation over all frames for an input segment. The re- sulted se gment-lev el representation is then fed into two fully connected layers to classify the speakers in the training set. Af- ter training, speaker embeddings are extracted from the 512- dimensional af fine component of the first fully connected layer . 2.3.3. Deep ResNet W e follow the deep ResNet system as described in [29, 3, 30], and we increase the widths (number of channels) of the residual blocks from { 16, 32, 64, 128 } to { 32, 64, 128, 256 } . The net- work architecture contains three main components: a front-end ResNet, a pooling layer, and a feed-forward network. The front- end ResNet transforms the ra w feature into a high-le vel abstract representation. The subsequent pooling layer outputs a single utterance-lev el representation. Specifically , means statistics are accumulated for each feature map, and finally 256-dimensional utterance-lev el representation is produced. Each unit in the out- put layer is represented as a target speak er identity . All the components in the pipeline are jointly learned in an end-to-end manner with a unified loss function. W e adopt the typical softmax loss as well as the angular softmax loss (A- softmax) [31]. A-softmax learns angularly discriminativ e fea- tures by generating an angular classication margin between em- beddings of different classes. The superiority of A-softmax has been sho wn in both face recognition [31], language recognition and speaker recognition [3]. After training, the 256-dimensional utterance-lev el speaker embedding is extracted after the penultimate layer of the neural network for the given utterance. In the testing stage, the full- length feature sequence is directly fed into the network, without any truncate or padding operation. Based on the deep ResNet framework, we in vestigate mul- tiple kinds of short-term spectral features and loss functions. Finally , we hav e four networks trained with different setups: • Mfbank-8k + Softmax: ResNet system trained on Mfbank-8k features with softmax loss. • Mfbank-16k + Softmax: ResNet system trained on Mfbank-16k features with softmax loss. • Mfbank-16k + A-softmax: ResNet system trained on Mfbank-16k features with A-softmax loss. • Gfbank + A-softmax. ResNet system trained on Gfbank- features with A-softmax loss. 2.4. Back-end modeling In back-end modeling, we either use cosine similarity based scoring, or Probabilistic Linear Discriminant Analysis (PLDA) based scoring. T able 1: Development subset r esults for the speaker reco gnition task of the VOiCES fr om a distance challenge (SN r epresents Score Normalization, devW r epresents whitening using de velopment subset) Front-end Back-end WPE SN Development subset Evaluation minC actC EER[%] minC actC EER[%] MFCC i-vector PLD A - √ 0.4935 0.6747 6.33 0.8037 0.8294 12.92 CORAL + devW + PLD A √ √ 0.4527 0.4703 6.12 0.6870 0.6891 11.89 PNCC i-vector PLD A - √ 0.5073 0.6745 6.12 0.6791 0.7803 10.18 CORAL + devW + PLD A √ - 0.4594 0.4697 5.29 0.6498 0.7152 10.09 x-vector CORAL + PLD A - √ 0.4018 0.4151 4.96 0.6377 0.6492 09.13 CORAL + PLD A √ - 0.3617 0.3688 4.52 0.5417 0.5544 07.54 Mfbank-8k ResNet + Softmax CORAL + devW + PLD A - - 0.4557 0.5246 5.41 0.6608 0.7128 10.92 CORAL + devW + PLD A √ - 0.3934 0.4611 4.59 0.5929 0.6424 09.75 Mfbank-16k ResNet + Softmax cosine similarity - - 0.3608 1 3.81 0.6262 1 08.75 cosine similarity √ - 0.3245 1 3.02 0.5507 1 07.91 Mfbank-16k ResNet + A-Softmax cosine similarity - - 0.2735 1 2.73 0.4156 1 05.84 cosine similarity √ - 0.2485 1 2.41 0.3668 1 05.58 Gfbank ResNet + A-Softmax cosine similarity - - 0.3065 1 3.52 0.4411 1 06.78 cosine similarity √ - 0.2680 1 3.14 0.4056 1 06.49 2.4.1. Cosine similarity W e use cosine similarity as a scoring method for the ResNet based systems. The scores of any gi ven enrollment-test pair are calculated as the cosine similarity of the two embeddings. 2.4.2. Gaussian PLDA W e use Correlation Alignment (CORAL) [32, 33] to align the distributions of out-of-domain and in-domain features in an un- supervised way by aligning second-order statistics, i.e., covari- ance. T o minimize the distance between the covariance of the out-of-domain and in-domain features, a linear transformation A to the original source features and the Frobenius norm is used as matrix distance metric: min A k C ˆ S − C T k 2 F = min A k A T C S A − C T k 2 F (1) where C S and C T are cov ariance matrix of the source-domain and target-domain features, C ˆ S is covariance of the transformed source features, and k · k 2 F denotes the matrix Frobenius norm. The embeddings after domain adaptation are whitened and unit-length normalized. The whitening transforms is estimated with either the training set or the dev elopment subset. The Gaussian PLDA model [34] with a full covariance residual noise term is trained on the speaker discriminant fea- tures. After the PLD A is trained, the scores of any giv en enrollment-test pair are calculated as the log-likelihood ratio on the PLD A model. 2.5. Score normalization After scoring, results from all trials are subject to score normal- ization. W e utilize Adaptive Symmetric Score Normalization (AS-Norm) in our systems [35]. The adaptiv e cohort for the en- rollment file are selected to be X closest (most positive scores) files to the enrollment utterance e as E top e . The cohort scores based on such selections for the enrollment utterance are then: S e ( E top e ) = { s ( e, ε ) |∀ ε ∈ E top e } (2) Then the AS-Norm is ˜ s ( e, t ) = 1 2  s ( e, t ) − µ [ S e ( E top e )] σ [ S e ( E top e )] + s ( e, t ) − µ [ S t ( E top t )] σ [ S t ( E top t )]  (3) 2.6. System fusion and calibration All the subsystems are fused and calibrated using the BOSARIS toolkit [36] which learn a scale and a bias for each subsystem. The final fusion is a score-lev el equal-weighted sum after ap- plying the scale and the bias. 3. Experiments 3.1. Data usage The training data includes V oxCeleb 1 [37] and V oxCeleb 2 [38]. The original distribution of V oxCeleb split each video into multiple short segments. During training, the segments from the same video are concatenated into a single sound wa ve, which results in 167897 utterances from 7245 speakers. No voice acti vity detection (V AD) is applied. For the dev elopment data, we only use a subset of the dev el- opment dataset provided by the V OiCES challenge. The total of 196 speakers in the original development dataset is split into two subgroups, each with 98 speakers. One subset is used as the new dev elopment set, and the other is used as the domain adap- tation and score normalization corpus. In this way , we reduce the original 4,005,888 trials into 999,424 trials. Since a part of the de velopment, data is used as the domain adaption and score normalization data, we can not pro vide the e xperimental results on the whole development data. So all the experimental results on the de velopment set pr esented in this paper use the new sub- trials. 3.2. System performance on single systems In table 1, the systems of different front-end speaker discrimi- nant features with the top one back-end are provided. 0.1 0.2 0.5 1 2 5 10 20 40 False Alarm probability (in %) 0.1 0.2 0.5 1 2 5 10 20 40 Miss probability (in %) Development original wav MFCC i-vector PNCC i-vector ResNet 8k x-vector ResNet 16k ResNet Gammatome ResNet A-softmax 0.1 0.2 0.5 1 2 5 10 20 40 False Alarm probability (in %) 0.1 0.2 0.5 1 2 5 10 20 40 Development dereverberated wav MFCC i-vector PNCC i-vector ResNet 8k x-vector ResNet 16k ResNet Gammatome ResNet A-softmax 0.1 0.2 0.5 1 2 5 10 20 40 60 80 90 False Alarm probability (in %) 0.1 0.2 0.5 1 2 5 10 20 40 60 Evaluation original wav MFCC i-vector PNCC i-vector ResNet 8k x-vector ResNet 16k ResNet Gammatome ResNet A-softmax 0.1 0.2 0.5 1 2 5 10 20 40 60 80 90 False Alarm probability (in %) 0.1 0.2 0.5 1 2 5 10 20 40 60 Evaluation dereverberated wav MFCC i-vector PNCC i-vector ResNet 8k x-vector ResNet 16k ResNet Gammatome ResNet A-softmax Figure 1: DET plots for development and evaluation dataset with original or der everberated sound wave T able 2: System performance on differ ent fusion system Fusion strategy Development subset Evaluation minC actC EER[%] Cllr minC actC EER[%] Cllr Best single system (ResNet + A-softmax + WPE) 0.2485 1 2.41 0.8060 0.3668 1 5.58 0.8284 Each embedding with top 1 back-end 0.1831 0.1857 1.93 0.0808 0.3205 0.3214 4.60 0.2335 Each embedding with top 2 back-end 0.1644 0.1659 1.48 0.0710 0.3555 0.3578 4.79 0.2684 Each embedding with top 3 back-end (submission) 0.1473 0.1484 1.21 0.0577 0.3532 0.3609 4.96 0.2683 From the results in table 1, several observations are drawn as follo ws. First, the PNCC based i-vector system obtains a no- ticeable performance gain under strong rev erberation and low SNR (signal to noise ratio) en vironments (ev aluation set) com- pared to MFCC based i-v ector system. For the dev elopment set with mild reverberation and higher SNR (about 20dB), the per- formance gain is not so obvious. Also, the WPE dere verberation algorithm results in 10% gain compared to the original wa ve for both i-vector and neural network based systems. Moreo ver , the ResNet + softmax system trained on 16k Mfbank achiev es 17.5% relativ e performance gain in terms of minDCF compared to the 8k Mfbank. The last observation from the results is the performance of the system with A-softmax loss. Compared to the ResNet + softmax system, the ResNet + A-softmax sys- tem significantly improv e t he system performance by more than 20% on both dev elopment and ev aluation sets. The Detection Error T radeoff (DET) curves in figure 1 pro- vide a clear comparison among the subsystems we used in the V OiCES challenge. The final best signal system is the ResNet + A-softmax network combined with cosine similarity scoring. Applying derev erberation to the enrollment and testing data can further improv e the performance. On the development set, the final minDCF and EER are 0.2485 and 2.41% respectiv ely . On the ev aluation set, the final minDCF and EER are 0.3668 and 5.58%. The performance degradation on the evaluation set can be observed from results. This performance degradation mainly due to the more challenging re verberation environments and much lower SNR in the ev aluation data, which lead to the mis- match between dev elopment and ev aluation data. 3.3. System performance on fused systems For the sev en kinds of front-end systems, the embeddings from the original audio and the de-reverberated audio are extracted respectiv ely , resulting in 14 types of front-end speaker discrim- inant features. Then, different back-end modeling methods, in- cluding cosine scoring, a different set of PLD A modeling, and different setting of score normalization, are applied to these features. For each speaker embedding, the top three back-end methods with the best performance on the particular embedding are selected, and finally , we get 42 indi vidual scores for the final fusion. The final results on the development subset and the ev al- uation set are shown in table 2. Our final submission obtains minDCF of 0.1473 and 0.3532 on the development and ev alua- tion set respectiv ely . After the ev aluation, we in vestigate the system performance fused with different back-ends. It is interesting to find that although fusion with the top 3 back-ends for each front-end embeddings improves the performance by 20% relatively com- pared to fusion with top 1 back-ends, the results on the evalua- tion show the opposite: fusion with the top 3 back-ends for each front-ends degrades the performance by 10% compared to the fused system with top 1 back-ends. This is mainly because of the mismatch between the dev elopment and ev aluation data. 4. Conclusions W e presented the components and analyzed the results of the DKU-SMIIP speaker recognition system for the VOiCES from a Distance Challenge 2019. W e use different acoustic fea- tures, different front-end modeling methods, and various back- end scoring methods. T o further improve the performance, we use WPE to dere verberate the dev elopment and ev aluation data. This enabled a series of incremental improvements, and the fu- sion showed that different subsystems are complementary to each other at score lev el. 5. References [1] N. Dehak, P . J. Kenny , R. Dehak, P . Dumouchel, and P . Ouellet, “Front- End F actor Analysis for Speaker V erification, ” IEEE Tr ansactions on Au- dio, Speech, and Language Pr ocessing , vol. 19, no. 4, pp. 788–798, 2011. [2] D. Snyder, D. Garcia-Romero, G. Sell, D. Povey , and S. Khudanpur, “x- vectors: Robust DNN Embeddings for Speaker Recognition, ” in IEEE In- ternational Conference on Acoustics, Speec h and Signal Processing , 2018, pp. 5329–5333. [3] W . Cai, J. Chen, and M. Li, “Exploring the Encoding Layer and Loss Function in End-to-End Speaker and Language Recognition System, ” in Odyssey: The Speaker and Langua ge Recognition W orkshop , 2018, pp. 74– 81. [4] M. W olfel and J. McDonough, Distant Speech Recognition . John Wile y & Sons, Incorporated, 2009. [5] P . Assmann and Q. Summerfield, “The Perception of Speech Under Ad- verse Conditions, ” in Speec h Pr ocessing in the A uditory System . Springer New Y ork, 2004, pp. 231–308. [6] T . Nakatani, T . Y oshioka, K. Kinoshita, M. Miyoshi, and Biing-Hwang Juang, “Speech Dereverberation Based on V ariance-Normalized Delayed Linear Prediction,” IEEE Tr ansactions on Audio, Speech, and Language Pr ocessing , vol. 18, no. 7, pp. 1717–1731, 2010. [7] X. Zhao, Y . W ang, and D. W ang, “Robust Speaker Identification in Noisy and Re verberant Conditions, ” IEEE/ACM Transactions on Audio, Speech, and Language Pr ocessing , vol. 22, no. 4, pp. 836–845, 2014. [8] M. K olboek, Z.-H. T an, and J. Jensen, “Speech Enhancement Using Long Short-T erm Memory based Recurrent Neural Networks for Noise Robust Speaker V erification, ” in IEEE Spoken Language T echnology W orkshop , 2016, pp. 305–311. [9] Z. Oo, Y . Kawakami, L. W ang, S. Nakaga wa, X. Xiao, and M. Iwahashi, “DNN-Based Amplitude and Phase Feature Enhancement for Noise Robust Speaker Identification, ” in Proceedings of the Annual Confer ence of the International Speech Communication Association , 2016, pp. 2204–2208. [10] S. E. Eskimez, P . Soufleris, Z. Duan, and W . Heinzelman, “Front-end speech enhancement for commercial speaker verification systems, ” Speech Communication , vol. 99, pp. 101–113, 2018. [11] J. Heymann, L. Drude, and R. Haeb-Umbach, “Neural Network Based Spectral Mask Estimation for Acoustic Beamforming, ” in 2016 IEEE Inter- national Conference on Acoustics, Speech and Signal Processing . IEEE, 2016, pp. 196–200. [12] E. W arsitz and R. Haeb-Umbach, “Blind Acoustic Beamforming Based on Generalized Eigenvalue Decomposition, ” IEEE Tr ansactions on Audio, Speech and Language Pr ocessing , vol. 15, no. 5, pp. 1529–1539, 2007. [13] T . Falk and W ai-Yip Chan, “Modulation Spectral Features for Robust Far- Field Speaker Identification, ” IEEE T ransactions on Audio, Speech, and Language Pr ocessing , vol. 18, no. 1, pp. 90–100, 2010. [14] S. O. Sadjadi and J. H. Hansen, “Hilbert Envelope Based Features for Ro- bust Speaker Identification Under Reverberant Mismatched Conditions, ” in 2011 IEEE International Conference on Acoustics, Speech and Signal Pro- cessing , 2011, pp. 5448–5451. [15] Q. Jin, R. Li, Q. Y ang, K. Laskowski, and T . Schultz, “Speaker Identifica- tion with Distant Microphone Speech, ” in 2010 IEEE International Confer- ence on Acoustics, Speech and Signal Pr ocessing , 2010, pp. 4518–4521. [16] S. O. Sadjadi and J. H. L. Hansen, “Blind Spectral W eighting for Robust Speaker Identification under Re verberation Mismatch, ” IEEE/ACM Tr ans- actions on Audio, Speech, and Language Pr ocessing , vol. 22, no. 5, pp. 937–945, 2014. [17] I. Peer, B. Rafaely , and Y . Zigel, “Reverberation Matching for Speaker Recognition, ” in 2008 IEEE International Conference on Acoustics, Speech and Signal Pr ocessing , 2008, pp. 4829–4832. [18] A. R. A vila, M. Sarria-Paja, F . J. Fraga, D. O’Shaughnessy , and T . H. Falk, “Improving the Performance of Far-Field Speaker V erification Using Multi- Condition Training: The Case of GMM-UBM and i-V ector Systems, ” in Pr oceedings of the Annual Conference of the International Speech Com- munication Association , 2014, pp. 1096–1100. [19] D. Garcia-Romero, X. Zhou, and C. Y . Espy-Wilson, “Multicondition train- ing of Gaussian PLDA models in i-vector space for noise and reverbera- tion robust speaker recognition, ” in 2012 IEEE International Conference on Acoustics, Speech and Signal Pr ocessing , 2012, pp. 4257–4260. [20] M. K. Nandwana, J. v an Hout, M. McLaren, A. Stauffer , C. Richey , A. Law- son, and M. Graciarena, “Robust Speaker Recognition from Distant Speech under Real Reverberant En vironments Using Speaker Embeddings, ” in Pr o- ceedings of the Annual Conference of the International Speech Communi- cation Association , 2018, pp. 1106–1110. [21] Q. Jin, T . Schultz, and A. W aibel, “Far-Field Speaker Recognition, ” IEEE T ransactions on A udio, Speech and Language Pr ocessing , vol. 15, no. 7, pp. 2023–2032, 2007. [22] Mikyong Ji, Sungtak Kim, Hoirin Kim, and Ho-Sub Y oon, “T ext- Independent Speaker Identification using Soft Channel Selection in Home Robot Environments, ” IEEE Tr ansactions on Consumer Electronics , vol. 54, no. 1, pp. 140–144, 2008. [23] M. K. Nandwana, J. V . Hout, M. McLaren, C. Richey , A. Lawson, and M. A. Barrios, “The VOiCES from a Distance Challenge 2019 Evaluation Plan, ” arXiv:1902.10828 [eess.AS] , 2019. [24] R. Scheibler, E. Bezzam, and I. Dokmanic, “Pyroomacoustics: A Python Package for Audio Room Simulation and Array Processing Algorithms, ” in 2018 IEEE International Conference on Acoustics, Speech and Signal Pr ocessing , 2018, pp. 351–355. [25] C. Richey , M. A. Barrios, Z. Armstrong, C. Bartels, H. Franco, M. Gra- ciarena, A. Lawson, M. K. Nandwana, A. Stauffer , J. van Hout, P . Gam- ble, J. Hetherly , C. Stephenson, and K. Ni, “V oices Obscured in Complex En vironmental Settings (V OICES) corpus, ” in Proceedings of the Annual Confer ence of the International Speec h Communication Association , 2018, pp. 1566–1570. [26] D. Snyder , G. Chen, and D. Po vey , “MUSAN: A Music, Speech, and Noise Corpus, ” arXiv:1510.08484 [cs] , 2015. [27] C. Kim and R. M. Stern, “Power-Normalized Cepstral Coefcients (PNCC) for Robust Speech Recognition, ” IEEE/ACM T ransactions on Audio, Speech and Language Pr ocessing , vol. 24, no. 7, pp. 1315–1329, 2016. [28] R. Patterson, K. Robinson, J. Holdsworth, D. McKeown, C. Zhang, and M. Allerhand, “Comple x Sounds and Auditory Images, ” in A uditory Physi- ology and P erception . Oxford, UK: Y . Cazals, L. Demany , and K. Horner, (Eds), Pergamon Press, 1992, pp. 429–446. [29] W . Cai, Z. Cai, W . Liu, X. W ang, and M. Li, “Insights into End-to-End Learning Scheme for Language Identification, ” in 2018 IEEE International Confer ence on Acoustics, Speech and Signal Processing , 2018, pp. 5209– 5213. [30] W . Cai, J. Chen, and M. Li, “ Analysis of length normalization in end-to- end speaker verification system, ” in Proc. INTERSPEECH 2018 , 2018, pp. 3618–3622. [31] W . Liu, Y . W en, Z. Y u, M. Li, B. Raj, and L. Song, “Sphereface: Deep Hypersphere Embedding for Face Recognition, ” in The IEEE Confer ence on Computer V ision and P attern Recognition , 2017, pp. 212–220. [32] B. Sun, J. Feng, and K. Saenko, “Return of Frustratingly Easy Domain Adaptation, ” in Proceedings of the Thirtieth AAAI Conference on Artificial Intelligence , 2016, pp. 2058–2065. [33] M. J. Alam, G. Bhattacharya, and P . Kenny , “Speaker V erification in Mismatched Conditions with Frustratingly Easy Domain Adaptation, ” in Odyssey: The Speaker and Language Recognition W orkshop , 2018. [34] D. Garcia-Romero and C. Y . Espy-Wilson, “ Analysis of i-vector Length Normalization in Speak er Recognition Systems, ” in Pr oceedings of the An- nual Conference of the International Speech Communication Association , 2011, pp. 249–252. [35] P . Matjka, O. Nov otn, O. Plchot, L. Burget, M. D. Snchez, and J. ernock, “ Analysis of Score Normalization in Multilingual Speaker Recognition, ” in Pr oceedings of the Annual Confer ence of the International Speech Commu- nication Association , 2017. [36] N. Br ¨ ummer and E. De V illiers, “The BOSARIS T oolkit: Theory , Algorithms and Code for Surviving the New DCF, ” arXiv preprint arXiv:1304.2865 , 2013. [37] A. Nagrani, J. S. Chung, and A. Zisserman, “V oxceleb: A Large-Scale Speaker Identification Dataset, ” in Pr oceedings of the Annual Confer ence of the International Speec h Communication Association , 2017, pp. 2616– 2620. [38] J. S. Chung, A. Nagrani, and A. Zisserman, “V oxceleb2: Deep Speaker Recognition, ” in Proceedings of the Annual Conference of the International Speech Communication Association , 2018.