Joint Speaker Counting, Speech Recognition, and Speaker Identification for Overlapped Speech of Any Number of Speakers

by   Naoyuki Kanda, et al.

We propose an end-to-end speaker-attributed automatic speech recognition model that unifies speaker counting, speech recognition, and speaker identification on monaural overlapped speech. Our model is built on serialized output training (SOT) with attention-based encoder-decoder, a recently proposed method for recognizing overlapped speech comprising an arbitrary number of speakers. We extend SOT by introducing a speaker inventory as an auxiliary input to produce speaker labels as well as multi-speaker transcriptions. All model parameters are optimized by speaker-attributed maximum mutual information criterion, which represents a joint probability for overlapped speech recognition and speaker identification. Experiments on LibriSpeech corpus show that our proposed method achieves significantly better speaker-attributed word error rate than the baseline that separately performs overlapped speech recognition and speaker identification.



page 1

page 2

page 3

page 4


Minimum Bayes Risk Training for End-to-End Speaker-Attributed ASR

Recently, an end-to-end speaker-attributed automatic speech recognition ...

Transcribe-to-Diarize: Neural Speaker Diarization for Unlimited Number of Speakers using End-to-End Speaker-Attributed ASR

This paper presents Transcribe-to-Diarize, a new approach for neural spe...

End-to-End Speaker-Attributed ASR with Transformer

This paper presents our recent effort on end-to-end speaker-attributed a...

Serialized Output Training for End-to-End Overlapped Speech Recognition

This paper proposes serialized output training (SOT), a novel framework ...

Domain Adversarial Neural Networks for Dysarthric Speech Recognition

Speech recognition systems have improved dramatically over the last few ...

Monaural Multi-Talker Speech Recognition using Factorial Speech Processing Models

A Pascal challenge entitled monaural multi-talker speech recognition was...

Revisiting joint decoding based multi-talker speech recognition with DNN acoustic model

In typical multi-talker speech recognition systems, a neural network-bas...
This week in AI

Get the week's most popular data science and artificial intelligence research sent straight to your inbox every Saturday.

1 Introduction

Speaker-attributed automatic speech recognition (SA-ASR) from overlapped speech has been an active research area towards meeting transcription

[1, 2, 3]. It requires to count the number of speakers, transcribe utterances that are sometimes overlapped, and also diarize or identify the speaker of each utterance. While significant progress has been made especially for multi-microphone settings (e.g., [4]), SA-ASR remains very challenging when we can only access monaural audio.

A significant amount of research has been conducted to achieve the goal of SA-ASR. One approach is applying speech separation (e.g., [5, 6, 7]) before ASR and speaker diarization/identification. However, a speech separation module is often designed with a signal-level criterion, which is not necessarily optimal for succeeding modules. To overcome this suboptimality, researchers have investigated approaches for jointly modeling multiple modules. For example, there are a number of studies concerning joint modeling of speech separation and ASR (e.g., [8, 9, 10, 11, 12, 13]). Several methods were also proposed for integrating speaker identification and speech separation [14, 15, 16]. However, little research has yet been done to address SA-ASR by combining all these modules.

Only a limited number of studies have tackled the joint modeling of multi-speaker ASR and speaker diarization/identification. [17]

proposed to generate transcriptions of different speakers interleaved by speaker role tags to recognize two-speaker conversations based on a recurrent neural network transducer (RNN-T). Although promising results were shown for two-speaker conversation data, the method cannot deal with speech overlaps due to the monotonicity constraint of RNN-T. Furthermore, their method is difficult to be extended to an arbitrary number of speakers because the speaker role tag needs to be uniquely defined for each speaker (e.g., a doctor and a patient).


proposed a joint decoding framework for overlapped speech recognition and speaker diarization, in which speaker embedding estimation and target-speaker ASR were applied alternately. Although this method is extendable to many speakers in theory, it assumes that the speaker counting is conducted during the speaker embedding estimation process, which is challenging in practice.

In this paper, we propose a joint framework for SA-ASR that entails speaker counting, overlapped speech recognition, and speaker identification. Our model is built on serialized output training (SOT) [19] with attention-based encoder-decoder (AED) [20, 21, 22, 23], which was recently proposed for recognizing overlapped speech consisting of an arbitrary number of speakers. We extend the SOT model by introducing a speaker inventory as an auxiliary input to produce speaker labels as well as multi-speaker transcriptions. All model parameters are optimized by maximizing a joint probability for overlapped speech recognition and speaker identification. Our model can recognize overlapped speech of any number of speakers while identifying the speaker of each utterance among any number of speaker profiles. We show that the proposed model achieves significantly better speaker-attributed word error rate (SA-WER) over the model consisting of separate modules.

2 Overlapped Speech Recognition with Serialized Output Training

2.1 ASR based on Attention-based Encoder Decoder

Given input

, an AED model produces a posterior probability of output sequence

as follows. Firstly, an encoder converts the input sequence into a sequence, , of embeddings, i.e.,


Secondly, at each decoder step , the attention module outputs attention weight as



is a decoder state vector at

-th step, and is the context vector at the previous time step. Then, context vector for the current time step is generated as a weighted sum of the encoder embeddings as follows.


Finally, the output distribution for is estimated given the context vector and decoder state vector as follows:


Here, we are assuming that and have the same dimensionality. Variable

is the affine matrix of the final layer. Note that

normally consists of a single affine transform with a softmax output layer. However, it was found in [19] that inserting one LSTM just before the affine transform effectively improves the SOT model, so we follow that architecture.

2.2 Serialized Output Training

With the SOT framework, the references for multiple overlapped utterances are concatenated to form a single token sequence by inserting a special symbol representing a speaker change. For example, for the three-speaker case, the reference label will be given as , where represents -th token of -th utterance. Note that , a token for sequence end, is used only at the end of the entire sequence.

Because there are multiple permutations in the order of reference labels to form , some trick is needed to calculate the loss for AED. One simple yet effective approach in [19] is sorting the reference labels by their start times, which is called “first-in, first-out” (FIFO) training. This training scheme works with complexity of with respect to the number of speakers and outperforms a scheme that exhaustively considers all possible permutations [19]. In this paper, we always use this FIFO training scheme.

Unlike permutation invariant training [7, 8], in which the number of the output branches that a model has constrains the number of recognizable speakers, SOT has no such limitation. Refer to [19] for a detailed description of the SOT model.

3 Proposed method

3.1 Overview

Suppose that we have a speaker inventory , where is the number of speakers in the inventory and is a speaker profile vector (e.g., d-vector [24]) of the -th speaker. The goal of the proposed method is to estimate a serialized multi-speaker transcription accompanied by the speaker identity of each token given input and .

In this work, we assume that the profiles of all speakers involved in the input speech are included in . In other words, we assume there is no “unknown” speaker for speaker identification. As long as this condition holds, the speaker inventory may include any number of irrelevant speakers’ profiles. This is a typical setup in scheduled office meetings, where meeting organizers invite attendees whose voice profiles are pre-registered.

Figure 1: Proposed model.

3.2 Model Architecture

We start with the conventional AED represented by the blue blocks in Fig. 1. Firstly, we introduce one more encoder to represent the speaker characteristics of input as follows.


On top of this, for every decoder step , we apply the attention weight generated by the attention module of AED to extract attention-weighted vector of speaker embeddings .


Note that could be contaminated by interfering speech because some time frames include two or more speakers.

The speaker query RNN in Fig. 1 then generates a speaker query given the speaker embedding , previous output , and previous speaker query .


Based on the speaker query , an attention module for speaker inventory (shown as InventoryAttention in the diagram) estimates the attention weight for each profile in .


Here, we use the softmax function (Eq. (10

)) on the cosine similarity between the speaker query and speaker profile (Eq. (

9)), which was found to be the most efficient in our preliminary experiment. The attention weight can be seen as a posterior probability of speaker speaking the -th token given all previous tokens and speakers as well as and .


Finally, we calculate the attention-weighted speaker profile based on and as followings.


This weighted profile is appended to the input of . Specifically, we replace Eq. (5) by


where is a matrix to change the dimension of to that of . Terms and are obtained from Eq. (3) and (4), respectively. Note that the output distribution is now conditioned on and because of the addition of the weighted profile .

3.3 Training

During training, all the network parameters are optimized by maximizing as follows. We call it speaker-attributed maximum mutual information (SA-MMI) training.


From Eq. (14) to Eq. (15

), the chain rule is applied for

and alternately. Here, we introduce a scaling parameter to adjust the scale of the speaker estimation probability to that of ASR. Equation (16) shows that our training criterion can be factorized into two conditional probabilities defined in Eqs. (13) and (11), respectively. Note that the speaker identity of the token or is set the same as that of the preceding token.

3.4 Decoding

An extended beam search algorithm is used for decoding with the proposed method. In the conventional beam search for AED, each hypothesis contains estimated tokens accompanied by the posterior probability of the hypothesis. In addition to these, a hypothesis for the proposed method contains speaker estimation . Each hypothesis expands until is detected, and the estimated tokens in each hypothesis are grouped by to form multiple utterances. For each utterance, the average of values, including the last token corresponding to or , is calculated for each speaker. The speaker with the highest average score is selected as the predicted speaker of that utterance. Finally, when the same speaker is predicted for multiple utterances, those utterances are concatenated to form a single utterance.

ModelEval Set 1-speaker 2-speaker-mixed 3-speaker-mixed Total
Single-speaker ASR - 4.7 - - 66.9 - - 90.7 - - 68.4 -
SOT-ASR - 4.5 - - 10.3 - - 19.5 - - 13.9 -
SOT-ASR + random speaker assignment 87.4 4.5 175.2 82.8 23.4 169.7 76.1 39.1 165.1 80.2 28.1 168.3
SOT-ASR + d-vec speaker identification 0.4 4.5 4.8 6.4 10.3 16.5 13.1 19.5 31.7 8.7 13.9 22.2
Proposed Model
SOT-ASR + Spk-Enc + Inv-Attn 0.3 4.3 4.7 5.5 10.4 12.2 14.8 23.4 26.7 9.3 15.9 18.2
+ SpeakerQueryRNN 0.4 4.2 4.6 3.0 9.1 10.9 11.6 21.5 24.7 6.9 14.5 16.7
+ Weighted Profile () 0.2 4.2 4.5 2.5 8.7 9.9 10.2 20.2 23.1 6.0 13.7 15.6
Table 1: SER (%), WER (%), and SA-WER (%) for baseline systems and proposed method. The number of profiles per test audio was 8. Each profile was extracted by using 2 utterances (15 sec on average). For random speaker assignment experiment (3rd row), averages of 10 trials were computed. No LM was used in the evaluation.
Actual # of Speakers Estimated # of Speakers (%)
in Test Data 1 2 3 4
1 99.96 0.04 0.00 0.00
2 2.56 97.44 0.00 0.00
3 0.31 25.34 74.35 0.00
Table 2: Speaker counting accuracy (%) of the proposed model.
# of Profiles # of Speakers in Test Data
in Inventory 1 2 3 Total
4 0.1 / 4.5 1.8 / 9.4 8.8 / 22.3 5.0 / 15.0
8 0.2 / 4.5 2.5 / 9.9 10.2 / 23.1 6.0 / 15.6
16 0.8 / 5.1 2.9 / 10.6 11.3 / 23.8 6.8 / 16.3
32 0.9 / 5.4 4.6 / 11.9 11.6 / 24.0 7.5 / 16.9
Table 3: SER (%) / SA-WER (%) for different numbers of profiles in the speaker inventory.
# of Utterances # of Speakers in Test Data
per Profile 1 2 3 Total
1 0.9 / 5.6 3.8 / 11.5 11.2 / 24.8 7.0 / 17.2
2 0.2 / 4.5 2.5 / 9.9 10.2 / 23.1 6.0 / 15.6
5 0.04 / 4.2 2.1 / 9.5 9.7 / 22.6 5.6 / 15.2
10 0.08 / 4.3 2.0 / 9.4 9.5 / 22.3 5.4 / 15.0
Table 4: Impact of the number of profile extraction utterances on SER (%) and SA-WER (%) for 8-profile setting. The average utterance duration was 7.5 sec.

4 Experiments

4.1 Evaluation settings

4.1.1 Evaluation data

We evaluated the effectiveness of the proposed method by simulating multi-speaker signals based on the LibriSpeech corpus [25]. Following the Kaldi [26] recipe, we used the 960 hours of LibriSpeech training data (“train_960”) for model learning, the “dev_clean” set for adjusting hyper-parameter values, and the “test_clean” set for testing.

Our training data were generated as follows. For each utterance in train_960, randomly chosen train_960 utterances were added after being shifted by random delays. When mixing the audio signals, the original volume of each utterance was kept unchanged, resulting in an average signal-to-interference ratio of about 0 dB. As for the delay applied to each utterance, the delay values were randomly chosen under the constraints that (1) the start times of the individual utterances differed by 0.5 sec or longer and that (2) every utterance in each mixed audio sample had at least one speaker-overlapped region with other utterances. For each training sample, speaker profiles were generated as follows. First, the number of profiles was randomly selected from to 8. Among those profiles, profiles were for the speakers involved in the overlapped speech. The utterances for creating the profiles of these speakers were different from those constituting the input overlapped speech. The rest of the profiles were randomly extracted from the other speakers in train_960. Each profile was extracted by using 10 utterances. We generated data for and combined them to use for training.

The development and evaluation sets were generated from dev_clean or test_clean, respectively, in the same way as the training set except that constraint (1) was not imposed. Therefore, multiple utterances were allowed to start at the same time in evaluation. Also, each profile was extracted from 2 utterances (15 sec on average) instead of 10, unless otherwise stated.

4.1.2 Evaluation metric

We evaluated the model with respect to speaker error rate (SER), WER, and SA-WER. SER is defined as the total number of model-generated utterances with speaker misattribution divided by the number of reference utterances. All possible permutations of the hypothesized utterances were examined by ignoring the ASR results, and the one that yielded the smallest number of errors (including the speaker insertion and deletion errors) was picked for the SER calculation. Similarly, WER was calculated by picking the best permutation in terms of word errors (i.e., speaker labels were ignored). Finally, SA-WER was calculated by comparing the ASR hypothesis and the reference transcription of each speaker.

4.1.3 Model settings

In our experiments, we used a 80-dim log mel filterbank, extracted every 10 msec, for the input feature. We stacked 3 frames of features and applied the model on top of the stacked features. For the speaker profile, we used a 128-dim d-vector [24], whose extractor was separately trained on VoxCeleb Corpus [27, 28]. The d-vector extractor consisted of 17 convolution layers followed by an average pooling layer, which was a modified version of the one presented in [29].

Our AsrEncoder consisted of 5 layers of 1024-dim bidirectional long short-term memory (BLSTM), interleaved with layer normalization 

[30]. The DecoderRNN consisted of 2 layers of 1024-dim unidirectional LSTM, and the DecoderOut consisted of 1 layer of 1024-dim unidirectional LSTM. We used a conventional location-aware content-based attention [22] with a single attention head. The SpeakerEncoder had the same architecture as the d-vector extractor except for not having the final average pooling layer. Our SpeakerQueryRNN consisted of 1 layer of 512-dim unidirectional LSTM. We used 16k subwords based on a unigram language model [31] as a recognition unit. We appplied volume perturbation to the mixed audio to increase the training data variability. Note that we applied neither an additional language model (LM) nor any other forms of data augmentation [32, 33, 34, 35] for simplicity.

Model training was performed as follows. In our preliminary experiment, training models from fully random parameters showed poor convergence due to the difficulty in attention module training. Therefore, we initialized the parameters of AsrEncoder, Attention, DecoderRNN, and DecoderOut by SOT-ASR parameters trained on simulated mixtures of LibriSpeech utterances as reported in [19]. We pre-trained the SOT-model with 640k iterations. We also initialized the SpeakerEncoder parameters by using those of the d-vector extractor. After the initialization, we updated the entire network based on with by using an Adam optimizer with a learning rate of 0.00002. We used 8 GPUs, each of which worked on 6k frames of minibatch. We report the results of the dev_clean-based best models found after 160k of training iterations.

4.2 Evaluation results

4.2.1 Baseline results

We built 4 different baseline systems, whose results are shown in the first 4 rows of Table 1. The first row corresponds the conventional single-speaker ASR based on AED. As expected, the WER was significantly degraded for overlapped speech. The second row shows the result of the SOT-ASR system that was used for initializing the proposed method in training. SOT-ASR significantly improved the WER for all evaluation settings. The lower WER for the 1-speaker case could be attributed to the data augmentation effect resulting from the use of overlapped speech for training, which was also observed in [19].

The third row shows the result of randomly assigning a speaker label for each utterance generated by SOT-ASR. Note that the speaker identification may affect WER as well as SA-WER. This is because multiple SOT-ASR-generated utternaces were mereged when their speaker labels were the same.

The fourth row shows the result of combining SOT-ASR and d-vector based speaker identification. In this baseline system, for each utterance, we calculated a weighted average of frame-level d-vectors by using the attention weights from SOT-ASR. The estimated d-vectors were then compared with each profile contained in the speaker inventory in terms of cosine similarity. The best scored speaker was selected one-by-one with a constraint that the same speaker could not be selected for multiple utterances. This method gave us reasonable results as can be seen in the table although the SA-WERs were not sufficient for overlapped speech.

4.2.2 Results of the proposed method

The last 3 rows of Table 1 shows the results of the proposed method while the first two of them were the results of an ablation study. “SOT-ASR + Spk-Enc + Inv-Attn” is the result of a variant of the proposed model where output was directly used for (Eq. (9)) instead of using output , and was not used in Eq. (13). Due to SA-MMI training, even this model achieved a lower SA-WER than the baseline while the SER and WER were degraded. Then, the entire performance was significantly boosted by introducing as shown in the next row. Finally, by introducing the weighted profile in Eq. (13

), the proposed method outperformed the baseline in all three evaluation metrics, resulting in 29% reduction of the SA-WER.

Table 2 shows the speaker counting accuracy of the proposed method. We can see that the speakers were counted very accurately especially for the 1-speaker (99.96%) and 2-speaker cases (97.44%) while it sometimes underestimated the number of the speakers for the 3-speaker mixtures.

4.2.3 Evaluation with different profile settings

In the previous experiments, we used the inventory comprising 8 profiles, each of which was extracted from 2 utterances. We then evaluated the proposed method with different numbers of profiles. As shown in Table 3, our proposed method showed only minor degradation in terms of the SER and SA-WER even with 32 profiles. This demonstrates the robustness of our method against the increase of profiles.

Finally, we also evaluated the impact of the number of utterances used for speaker profile extraction. As shown in Table 4, using more utterances for a profile yielded lower error rates.

5 Conclusions

In this paper, we proposed a joint model for SA-ASR that can recognize overlapped speech of any number of speakers while identifying the speaker of each utterance among any number of speaker profiles. In the experiments on LibriSpeech, the proposed model achieved significantly better SA-WER than the baseline that consists of separated modules.


  • [1] J. G. Fiscus, J. Ajot, and J. S. Garofolo, “The rich transcription 2007 meeting recognition evaluation,” in Multimodal Technologies for Perception of Humans.   Springer, 2007, pp. 373–389.
  • [2] A. Janin, D. Baron, J. Edwards, D. Ellis, D. Gelbart, N. Morgan, B. Peskin, T. Pfau, E. Shriberg, A. Stolcke et al., “The ICSI meeting corpus,” in Proc. ICASSP, vol. 1, 2003, pp. I–I.
  • [3] J. Carletta, S. Ashby, S. Bourban, M. Flynn, M. Guillemot, T. Hain, J. Kadlec, V. Karaiskos, W. Kraaij, M. Kronenthal et al., “The AMI meeting corpus: A pre-announcement,” in

    International workshop on machine learning for multimodal interaction

    .   Springer, 2005, pp. 28–39.
  • [4] T. Yoshioka, I. Abramovski, C. Aksoylar, Z. Chen, M. David, D. Dimitriadis, Y. Gong, I. Gurvich, X. Huang, Y. Huang et al., “Advances in online audio-visual meeting transcription,” in Proc. ASRU, 2019, pp. 276–283.
  • [5] J. R. Hershey, Z. Chen, J. Le Roux, and S. Watanabe, “Deep clustering: Discriminative embeddings for segmentation and separation,” in Proc. ICASSP, 2016, pp. 31–35.
  • [6] Z. Chen, Y. Luo, and N. Mesgarani, “Deep attractor network for single-microphone speaker separation,” in Proc. ICASSP, 2017, pp. 246–250.
  • [7] D. Yu, M. Kolbæk, Z.-H. Tan, and J. Jensen, “Permutation invariant training of deep models for speaker-independent multi-talker speech separation,” in Proc. ICASSP.   IEEE, 2017, pp. 241–245.
  • [8] D. Yu, X. Chang, and Y. Qian, “Recognizing multi-talker speech with permutation invariant training,” Proc. Interspeech 2017, pp. 2456–2460, 2017.
  • [9] H. Seki, T. Hori, S. Watanabe, J. Le Roux, and J. R. Hershey, “A purely end-to-end system for multi-speaker speech recognition,” in Proc. ACL, 2018, pp. 2620–2630.
  • [10] X. Chang, Y. Qian, K. Yu, and S. Watanabe, “End-to-end monaural multi-speaker ASR system without pretraining,” in Proc. ICASSP, 2019, pp. 6256–6260.
  • [11] X. Chang, W. Zhang, Y. Qian, J. L. Roux, and S. Watanabe, “MIMO-SPEECH: End-to-end multi-channel multi-speaker speech recognition,” in Proc. ASRU, 2019, pp. 237–244.
  • [12] N. Kanda, Y. Fujita, S. Horiguchi, R. Ikeshita, K. Nagamatsu, and S. Watanabe, “Acoustic modeling for distant multi-talker speech recognition with single-and multi-channel branches,” in Proc. ICASSP, 2019, pp. 6630–6634.
  • [13] N. Kanda, S. Horiguchi, R. Takashima, Y. Fujita, K. Nagamatsu, and S. Watanabe, “Auxiliary interference speaker loss for target-speaker speech recognition,” in Proc. Interspeech, 2019, pp. 236–240.
  • [14] P. Wang, Z. Chen, X. Xiao, Z. Meng, T. Yoshioka, T. Zhou, L. Lu, and J. Li, “Speech separation using speaker inventory,” in Proc. ASRU, 2019, pp. 230–236.
  • [15] T. von Neumann, K. Kinoshita, M. Delcroix, S. Araki, T. Nakatani, and R. Haeb-Umbach, “All-neural online source separation, counting, and diarization for meeting analysis,” in Proc. ICASSP, 2019, pp. 91–95.
  • [16] K. Kinoshita, M. Delcroix, S. Araki, and T. Nakatani, “Tackling real noisy reverberant meetings with all-neural source separation, counting, and diarization system,” arXiv preprint arXiv:2003.03987, 2020.
  • [17] L. El Shafey, H. Soltau, and I. Shafran, “Joint speech recognition and speaker diarization via sequence transduction,” in Proc. Interspeech, 2019, pp. 396–400.
  • [18] N. Kanda, S. Horiguchi, Y. Fujita, Y. Xue, K. Nagamatsu, and S. Watanabe, “Simultaneous speech recognition and speaker diarization for monaural dialogue recordings with target-speaker acoustic models,” in Proc. ASRU, 2019.
  • [19] N. Kanda, Y. Gaur, X. Wang, Z. Meng, and T. Yoshioka, “Serialized output training for end-to-end overlapped speech recognition,” arXiv preprint arXiv:2003.12687, 2020.
  • [20] D. Bahdanau, K. Cho, and Y. Bengio, “Neural machine translation by jointly learning to align and translate,” arXiv preprint arXiv:1409.0473, 2014.
  • [21] J. Chorowski, D. Bahdanau, K. Cho, and Y. Bengio, “End-to-end continuous speech recognition using attention-based recurrent NN: First results,” in

    NIPS Workshop on Deep Learning

    , 2014.
  • [22] J. K. Chorowski, D. Bahdanau, D. Serdyuk, K. Cho, and Y. Bengio, “Attention-based models for speech recognition,” in Proc. NIPS, 2015, pp. 577–585.
  • [23] W. Chan, N. Jaitly, Q. Le, and O. Vinyals, “Listen, attend and spell: A neural network for large vocabulary conversational speech recognition,” in Proc. ICASSP, 2016, pp. 4960–4964.
  • [24] E. Variani, X. Lei, E. McDermott, I. L. Moreno, and J. Gonzalez-Dominguez, “Deep neural networks for small footprint text-dependent speaker verification,” in Proc. ICASSP, 2014, pp. 4052–4056.
  • [25] V. Panayotov, G. Chen, D. Povey, and S. Khudanpur, “Librispeech: an ASR corpus based on public domain audio books,” in Proc. ICASSP, 2015, pp. 5206–5210.
  • [26] D. Povey, A. Ghoshal, G. Boulianne, L. Burget, O. Glembek, N. Goel, M. Hannemann, P. Motlicek, Y. Qian, P. Schwarz et al., “The Kaldi speech recognition toolkit,” in ASRU, 2011.
  • [27] A. Nagrani, J. S. Chung, and A. Zisserman, “Voxceleb: A large-scale speaker identification dataset,” in Proc. Interspeech, 2017, pp. 2616–2620.
  • [28] J. S. Chung, A. Nagrani, and A. Zisserman, “Voxceleb2: Deep speaker recognition,” in Proc. Interspeech, 2018, pp. 1086–1090.
  • [29] T. Zhou, Y. Zhao, J. Li, Y. Gong, and J. Wu, “CNN with phonetic attention for text-independent speaker verification,” in Proc. ASRU, 2019, pp. 718–725.
  • [30] J. L. Ba, J. R. Kiros, and G. E. Hinton, “Layer normalization,” arXiv preprint arXiv:1607.06450, 2016.
  • [31] T. Kudo, “Subword regularization: Improving neural network translation models with multiple subword candidates,” arXiv preprint arXiv:1804.10959, 2018.
  • [32] N. Kanda, R. Takeda, and Y. Obuchi, “Elastic spectral distortion for low resource speech recognition with deep neural networks,” in Proc. ASRU, 2013, pp. 309–314.
  • [33] T. Ko, V. Peddinti, D. Povey, and S. Khudanpur, “Audio augmentation for speech recognition,” in Proc. Interspeech, 2015, pp. 3586–3589.
  • [34] D. S. Park, W. Chan, Y. Zhang, C.-C. Chiu, B. Zoph, E. D. Cubuk, and Q. V. Le, “Specaugment: A simple data augmentation method for automatic speech recognition,” in Proc. Interspeech, 2019, pp. 2613–2617.
  • [35] C. Wang, Y. Wu, Y. Du, J. Li, S. Liu, L. Lu, S. Ren, G. Ye, S. Zhao, and M. Zhou, “Semantic mask for transformer based end-to-end speech recognition,” arXiv preprint arXiv:1912.03010, 2019.