Continuous speech separation: dataset and analysis

01/30/2020 ∙ by Zhuo Chen, et al. ∙ 0

This paper describes a dataset and protocols for evaluating continuous speech separation algorithms. Most prior studies on speech separation use pre-segmented signals of artificially mixed speech utterances which are mostly fully overlapped, and the algorithms are evaluated based on signal-to-distortion ratio or similar performance metrics. However, in natural conversations, a speech signal is continuous, containing both overlapped and overlap-free components. In addition, the signal-based metrics have very weak correlations with automatic speech recognition (ASR) accuracy. We think that not only does this make it hard to assess the practical relevance of the tested algorithms, it also hinders researchers from developing systems that can be readily applied to real scenarios. In this paper, we define continuous speech separation (CSS) as a task of generating a set of non-overlapped speech signals from a continuous audio stream that contains multiple utterances that are partially overlapped by a varying degree. A new real recorded dataset, called LibriCSS, is derived from LibriSpeech by concatenating the corpus utterances to simulate a conversation and capturing the audio replays with far-field microphones. A Kaldi-based ASR evaluation protocol is also established by using a well-trained multi-conditional acoustic model. By using this dataset, several aspects of a recently proposed speaker-independent CSS algorithm are investigated. The dataset and evaluation scripts are available to facilitate the research in this direction.



There are no comments yet.


page 1

page 2

page 3

page 4

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

As a natural phenomenon in human interactions, overlapping speech occupies a significant part of conversation time. This poses challenges for many speech technologies including automatic speech recognition (ASR) and speaker diarization because they usually assume one or zero speaker to be active at the same time. Speech separation could provide a solution for this problem.

The speech separation technology has been significantly improved over the past five years by leveraging deep learning. One fundamental challenge in overlapped speech separation is the inherent indeterminacy of the speaker order, which complicates supervised model training. To deal with this problem,

[11] proposed deep clustering (DC), achieving high quality single-channel speech separation for the first time by using a recurrent network with an affinity-based objective function that is invariant to the number of speakers and their permutation. In [37], the authors proposed permutation invariant training (PIT), which was shown to achieve a similar level of separation performance by exhaustively searching for the best speaker permutation during model training. Numerous extentions to these methods were proposed with different focuses [29, 33, 7, 6, 16, 27].

Another approach to the permutation indeterminacy problem is “informed extraction”, which makes use of additional information to distinguish a target speaker from other participating speakers. Use of visual information [8, 38], audio snippets of the speakers [40, 28, 32], or their locations [5, 22, 39] were investigated. On top of, or aside from, these algorithmic improvements, researchers also sought more effective input features [30, 36] and model architectures [18, 20, 24].

The signal to distortion ratio (SDR) [9]

and the scale-invariant signal-to-noise ratio (SISNR)

[15] have been steadily increasing on WSJ0-2mix [11], the most widely used speech separation dataset, which indicates the consistent progress of the separation technology. An early system [11] achieved SDR improvement of 6.3 dB while [19] improved the SDR by 19.0 dB. [20] reported that, in WSJ0-2mix, separated speech signals generated by TasNet, one of the state-of-the-art separation methods, were almost indistinguishable from clean utterances.

Despite those advances, existing speech separation evaluation schemes have several shortcomings that make it difficult to assess the practical relevance of the tested algorithms. Firstly, while most separation studies focus on disentagling fully overlapped speech signals, it is crucial that separation algorithms do not introduce signal distortion when one person is talking. An overlap ratio is below 20% in a natural meeting [4]. Therefore, the evaluation of separation algorithms has to be done in a way that considers both the separation accuracy for the overlapping periods and the distortionlessness for the overlap-free segments, which was not considered sufficiently before.

This also leads to the second issue. Most existing speech separation evaluation schemes use pre-segmented samples by implicitly assuming that an accurate overlap detector to be available. Some methods further assume prior knowledge of the number of participating speakers or knowledge of a target speaker. However, in practice, obtaining such information from conversational recordings is challenging by itself. In addition, some of these elements must be interconnected by nature. For example, accurate speaker counting would benefit from separating each person from the mixture. Focusing only on the separation accuracy could make us blind to this correlation between different elements.

In addition, most speech separation methods have been evaluated in terms of signal-based metrics such as SDR or SISNR. However, it is known that the signal-level performance metrics have weak correlations with ASR accuracy or perceptual sound quality.

To bring the speech separation technology to a real meeting transcription task, [36] proposed continuous speech separation (CSS), i.e., generating multiple overlap-free signals from an input audio stream that occasionally contains overlapped utterances. In this paper, we create a new dataset, called LibriCSS, recorded by using LibriSpeech utterances to facilitate CSS research while attempting to keep the task simple. Based on this dataset, we explore different aspects of the CSS method of [36]. We have made the dataset available along with Kaldi-based ASR evaluation scripts.

2 Continuous speech separation

In the most general way, continuous speech separation, or CSS, can be defined as a process of generating a set of overlap-free speech signals from a continuous audio stream containing multiple utterances spoken by different people. The original utterances contain both overlapped and overlap-free parts. In conversation scenarios like meetings, adjacent utterances overlap only by 10–15% [4] on average. When used as a front-end module, CSS allows downstream speech applications, such as ASR or speaker diarization, to operate on the assumption that only one person is active at each time point. Note that this task implicitly includes overlap detection as a subtask.

There are multiple approaches to CSS. One approach could be to keep generating an enhanced signal in a continuous fashion for each involved speaker as in [12, 25]

, which requires online speaker diarization. Alternatively, this could be achieved with offline processing by first performing speaker diarization and then extracting individual speaker signals by using an informed extraction approach based on the speaker embedding vectors 

[14]. The detect-then-separate strategy, i.e., performing overlap detection prior to separation, is an approach that many previous studies implicitly assumed. One potential problem with this approach is that separated signals for the overlapped segments need to be concatenated with the preceding and following signals in a way that keeps speaker consistency, which is not a trivial task. Also, overlap detection and speech separation are essentially inter-dependent problems. Thus, the sequential approach would lead to a suboptimal solution.

In this paper, we investigate the speaker-independent CSS approach proposed in [36]. With this approach, we output a fixed number () of audio streams, where each stream contains at most one active speaker at any time. For segments with no speaker overlaps, this CSS algorithm routes the incoming speech into one of the output channels, while the other output channels produce zero or negligible noise. The method was applied to real meeting recordings, where was set to 2 because three-fold overlaps rarely happen in meetings. It was shown to yield significant ASR accuracy gains compared with conventional beamformers for real meetings [35].

3 LibriCSS

Figure 1: Example recording setup (left) and microphone array geometry (right).

3.1 Dataset

The LibriCSS dataset is aimed at facilitating speech separation algorithm evaluation in the continuous input setting and thereby bridge the gap between the state of the speech separation research and what is required in real world applications. At the same time, the dataset is designed to be simple to make it broadly accessible. It consists of multi-channel audio recordings of “simulated conversations”, each containing multiple utterances spoken by different speakers. Each utterance is taken from LibriSpeech and played back from a loudspeaker placed in a room. We refer to each simulated conversation as a session in the following.

The dataset is designed to capture three features that are mostly missing in existing popular corpora. Firstly, the data are recorded in a room instead of being generated by simulation. The simulated data tend to oversimplify room acoustics especially in multi-channel scenarios. Secondly, the dataset encompasses different overlap ratios and silence settings to help analyze how different algorithms work under various overlap conditions. As discussed in Section 1, many existing studies only consider separating sufficiently overlapped utterances, overlooking the possibility of introducing signal distortion in non-overlapped segments. Thirdly, the audio signals are continuously recorded to enable CSS evaluation. Meanwhile, ground-truth segmentatations are also provided, allowing for the conventional utterance-wise evaluation.

LibriCSS consists of 10 hours of audio recordings. 10 sessions are included in the dataset, where each session is approximately one hour long. Each session is made up of six 10-minute-long “mini sessions” that have different overlap ratios (OVRs), ranging from 0 to 40%, where , with and being the total length of the overlapped speech and the total speech length, respectively. Each mini session includes eight speakers that are randomly selected from 40 speakers in the LibriSpeech development set. The total number of utterances in each mini session ranges from 52 to 125. For systems that do not perform speaker diarization before ASR, quick turn-taking is likely to result in concatenating multiple speaker utterances. This could end up with carrying over the internal state of a preceding speaker to the next speaker while decoding, which might degrade the ASR performance. To investigate this, for the 0% overlap scenario, two conditions are considered with respect to the length of silence between utterances. In both conditions, the utterances are played back sequentially without overlaps. In the short silence version, the inter-utterance silence length is randomly sampled from 0.1 to 0.5 seconds. The long silence version uses an inter-utterance silence period of 2.9–3.0 seconds.

The recordings were made in a regular meeting by using a seven-channel circular microphone array (see Fig. 1). The loudspeaker locations were randomly chosen in the meeting room while they remained the same for each session. The distances between the loudspeakers and the microphone array ranged from 33 cm to 409 cm.

3.2 Related Work

With the awareness of limitations in the existing speech separation evaluation schemes, several datasets were recently used other than WSJ0-2mix, focusing on different aspects of the speech separation. In [31], the authors collected replayed WSJ0 utterances in different outdoor environments to evaluate the noise robustness of speech separation systems. ASR evaluation was not considered. [1] used artificially simulated multi-channel WSJ0 utterances. While the paper reported word error rates (WERs), it used an extremely simple ASR system trained on clean utterances, making it hard to assess the usefulness of the separation algorithms. In addition, both datasets were collected and evaluated in an utterance-wise manner, with oracle utterance segmentation.

There are several corpora consisting of conversational recordings including natural speech overlaps. In [2, 26], “dinner party” style data were collected. However, systems were allowed to make use of oracle speaker segmentations. The evaluation was also performed in an utterance-wise fashion. Meetings are other situations where overlaps naturally happen [13, 3]. The meeting corpora may be used for evaluationg the speech separation algorithms in an end-to-end fashion as in [35], it is desirable to have a dataset that focuses on speech separation evaluation and allows for detailed analysis.

The LibriCSS dataset proposed in this work were collected under the same setup as [35]. The full training and ASR evaluation setup will be provided. This will allow speech separation algorithms to be easily tested in a more practical setting than [31, 1] while keeping the evaluation scheme simple.

3.3 Evaluation protocol

3.3.1 ASR setup

We use an ASR system to measure the speech separation accuracy. Our acoustic model was trained on 960 hours of LibriSpeech training data, which contains both clean and noisy audio. Kaldi [23]

was used to generate a phonetic decision tree and alignments. A bidirectional long short term memory (BLSTM) acoustic model was built using PyKaldi2 


– a toolbox that is developed on top of Kaldi and PyTorch. The BLSTM has 3 layers, and each layer has 512 cells for both forward and backward directions. We first trained the model using the cross-entropy criterion, and then fine tuned the model using the maximum mutual information (MMI). We used the standard 4-gram langauge model in LibriSpeech for decoding. Refer to 

[17] for more details about the ASR setup.

Two evaluation configurations are considered: one uses pre-segmented audio and one uses unsegmented audio. These are referred to as utterance-wise evaluation and continuous input evaluation, respectively. The latter is used for evaluating CSS algorithms.

3.3.2 Utterance-wise evaluation

In the utterance-wise evaluation, we firstly aligned the far-field meeting recording with the close talk reference using cross correlation, then each utterance in far-field recording was segmented using the ground truth utterance boundary from close talk. The separation was performed on the each utterance individually. The word error rate(WER) was applied as the evaluation metric. As two separated results were generated for each utterance, the result with lower WER were picked as the final result.

3.3.3 Continuous input evaluation

In the continuous input evaluation, the separation and recognition is designed to performed across utterance, i.e. no ground truth boundary information was introduced in each segments. However, as the online decoding is currently not well supported for Kaldi-based AM, feeding the whole meeting to the recognizer results in AM crash. To compensate this, we perform the segment-wise decoding as continuous evaluation instead, where we pre-segment the meetings into long segments that is around 60 120s in duration, based on the ground truth utterance boundary. More specifically, we ensure the boundary of segments is on the silence region in meeting. After segmentation, each segments contains around 8 10 utterance.

Then continuous speech separation was performed on each segments, resulting two continuous outputting streams. The stitching step discussed in section 2

was performed to estimate the final masks for each stream.

For each channel, We further segment each outputting channel based on silence to avoid the AM memory leakage between utterances. This step is relative easy as long silence can be observed for each outputting channel. A voice activity detector*** was applied on each separated channel with the mode “0”, i.e. the least aggressive non-speech filtering. The same procedure on channel 0 from original recording was performed, serving as the baseline for continuous evaluation.

Asclite tool was applied to generate speaker agnostic word error rate for continuous evaluation, which aligns multiple(two for our work) hypotheses against multiple speaker-specific reference transcriptions to generate word error rate (WER) estimates.

For both evaluation, session 0 in LibriCSS dataset serves as the development set for speech recognizer hyper-parameter searching. The recognition result for both beamforming and masking were reported.

4 Experimental Results

4.1 Continuous speech separation method

Figure 2: Generating TF masks in a streaming fashion by using a sliding window. Output orders for each pair of neighboring chunks are aligned by using the frames that the two chunks share.

The speaker-independent CSS algorithm proposed in [34] is implemented and evaluated. This would provide a reference point for calibrating the future results obtained on this dataset. The method uses a bidirectional model to estimate three time-frequency (TF) masks: two for speech sources and one for noise. To achieve this with streaming processing, a sliding window-based approach is employed as illustrated in Fig. 2. The window comprises three subwindows, each representing a past, current, or future context. They consist of , , and frames, respectively. At each operation point, the input features within the whole window are fed to the separation model to generate the TF masks. Then, only the TF masks within the current subwindow are used. The past and future subwindows help improve the mask estimation accuracy by providing left and right acoustic contexts to the separation model. The window is then shifted by

frames to process the next chunk of frames. Given the three TF masks, we output two speech signals with mask-based adaptive minimum variance distortionless response (MVDR) beamforming, following


Our separation model consists of three 1024-cell BLSTM layers (512 cells for each direction), followed by three parallel ReLU projection layers for mask estimation. The model is trained on 206 hours of artificially reverberated and mixed speech signals to minimize the Euclidean distance between reference and masked speech signals. The clean signals are randomly sampled from

train-clean-{100,360} [21]. For each mixture sample, room impulse responses are generated with the image method by assuming a random room with . Following [10], multi-channel isotropic noise signals are generated and added to the mixture signal at an SNR in the range of 0 to 10 dB. The training data configuration and simulated room impulse response will be released.

Four additional models are also built to examine the effect of the number and arrangement of microphones. One is based only on a single channel input. Two models are based on three-channel input: one uses microphones 1, 0, and 4 (i.e., a linear array) in Fig. 1; one uses microphones 1, 3, and 5 (i.e., a triangular array). The fourth model uses microphones 0, 1, 2, ,4, and 5.

4.2 Utterance-wise evaluation

Table 1 shows WERs for utterance wise evaluation. We can see that even a small amount of overlap severely degraded the ASR performance. While the single-channel separation model improved the WERs when the OVR was 10% or higher, it degraded the WER when there was not speech overlap, which was not sufficiently considered in the prior studies as we discussed earlier. The seven-channel model clearly generated more accurate TF masks outperforming the single-channel model in all conditions. However, when TF masking was performed, it did not improve the ASR accuracy for the 0% OVR case. When seven microphones were available, the MVDR performance significantly surpassed that of the TF masking. This is in alignment with the observation in [36]. While the MVDR-based seven-channel system provided a substantial WER gain, the performance difference between the 0% and 40% overlap cases was still significant, calling for the development of more accurate separation algorithms.


System Overlap ratio in %
0S 0L 10 20 30 40
No separation 11.8 11.7 18.8 27.2 35.6 43.3
Mask (1ch) 12.7 12.1 17.6 23.2 30.5 35.6
Mask (7ch) 12.0 11.6 15.6 20.2 25.6 29.4
MVDR (7ch) 8.4 8.3 11.6 15.8 18.7 21.7
Table 1: %WERs for utterance-wise evaluation. 0S: overlap ratio 0f 0% with short inter-utterance silence. 0L: overlap ratio of 0% with long inter-utterance silence. Our ASR system yields WERs of 4.9% and 5.1% for anechoic versions of 0S and 0L utterances.

4.3 Continuous input evaluation


System Overlap ratio in %
0S 0L 10 20 30 40
No separation 15.4 11.5 21.7 27.0 34.3 40.5
1.2-0.8-0.4 (1.2 s) 11.9 9.7 13.6 15.0 19.9 21.9
1.6-0.8-0.0 (0.8 s) 12.2 9.7 14.7 16.1 20.5 23.1
0.8-0.4-0.4 (0.8 s) 11.5 9.5 13.4 15.8 19.7 21.2
Table 2: %WERs for continuous input evaluation with seven microphones. Different chunking parameteres are examined. The dash-separated three numbers of the first column are , , values, respectively. The duration in parentheses represents inherent latency.


System Overlap ratio in %
0S 0L 10 20 30 40
7ch 11.9 9.7 13.6 15.0 19.9 21.9
5ch 12.8 10.5 15.3 17.4 22.8 26.4
3ch (triangular) 15.8 10.8 19.4 23.1 28.9 36.0
3ch (linear) 15.1 9.8 17.7 20.6 30.1 29.6
1ch 17.6 16.3 20.9 26.1 32.6 36.1
Table 3: %WER impact of number and arrangement of microphones in continuous input case. , , and are set at equivalents of 1.2 s, 0.8 s, and 0.4 s, respectively.

Table 2 shows the WERs for the continuous input evaluation with seven-channel input. It can be seen that the CSS algorithm improved the WERs for all conditions, including the cases where there was no speech overlap (i.e., 0S and 0L) thanks to the beamforming processing. The first system, denoted as 1.2-0.8-0.4, uses the past, current, and future acoustic contexts of 1.2 s (), 0.8 s (), and 0.4 s () , respectively, for mask estimation. This is the same configuration as the one used in [35]. This system has an inherent latency of 1.2 s (i.e., ). The inherent latency means the amount of delay caused by the system configuration, which does not include the processing delay resulting from actual computation for network evaluation, MVDR computation, and so on. In the last two rows of Table 2, two different configurations are examined, both of which reduces the inherent latency to 0.8 s. One way (1.6-0.8-0.0) is to avoid look-ahead by setting at 0. The chunk size i.e., , is kept constant. The other way (0.8-0.4-0.4) is to reduce the chunk size while keeping the look-ahead size constant. The results clearly show the usefulness of take account of the future acoustic context (i.e., keep at the equivalent of 0.4 s).

We can also see the detrimental effect of turn-taking without much silence. In the baseline system (i.e., without separation processing), the WER was increased by 33.9% relative just by reducing the inter-utterance gap from around 3 s (0L) to 0.5 s or shorter (0S) even though their WERs were almost the same when correct segmentations were provided (see Table 1). The CSS processing also mitigated this degradation, reducing the relative WER increase to 22.7%.

Table 3 lists the WERs for different microphone setups. The results clearly show the significant impact that the number of microphones has on the performance. No meaningful improvement was observed for the non-overlapping setting when three microphones were used. As with the utterance-wise evaluation results, the single-channel system even degraded the WERs for this setting while it still slightly provided gains for the other settings.

5 Conclusion

This paper described a dataset and protocols for evaluating continuous speech separation algorithms. The dataset is called LibriCSS and consists of multi-channel recordings of LibriSpeech utterances contatenated and replayed in a meeting room. By using a PIT-based speaker-independent CSS method, several aspects of CSS are investigated on this dataset.

Our experimental results shed light on the areas that need further improvement. Firstly, even with the seven-channel MVDR system, the performance degradation caused by speech overlap is not trivial. Also, in the single- and three-microphone cases, the separation processing does not improve or sometimes degrades the WER when only one speaker is active. We hope that this dataset and the associated evaluation pipeline facilitate speech sepration research while helping reduce the gap between the research state and what is required in real conversational processing applications.

6 Acknowledgement

We would like to thank Xiong Xiao for meaningful discussion and help in preparing source data.


  • [1] F. Bahmaninezhad, J. Wu, et al. (2019) A comprehensive study of speech separation: spectrogram vs waveform separation. arXiv preprint arXiv:1905.07497. Cited by: §3.2, §3.2.
  • [2] J. Barker, S. Watanabe, et al. (2018) The fifth’chime’speech separation and recognition challenge: dataset, task and baselines. arXiv preprint arXiv:1803.10609. Cited by: §3.2.
  • [3] J. Carletta, S. Ashby, et al. (2006) The ami meeting corpus: a pre-announcement. In Machine Learning for Multimodal Interaction, Berlin, Heidelberg, pp. 28–39. External Links: ISBN 978-3-540-32550-5 Cited by: §3.2.
  • [4] O. Çetin and E. Shriberg (2006) Analysis of overlaps in meetings by dialog factors, hot spots, speakers, and collection site: insights for automatic speech recognition. In Proc. Interspeech, pp. 293–296. Cited by: §1, §2.
  • [5] Z. Chen, X. Xiao, T. Yoshioka, et al. (2018) Multi-channel multi-speaker overlapped speech recognition with location guided speech extraction network. In Proc. SLT, Cited by: §1.
  • [6] Z. Chen and J. Droppo (2018) Sequence modeling in unsupervised single-channel overlapped speech recognition. In Proc. ICASSP 2018, pp. 4809–4813. Cited by: §1.
  • [7] Z. Chen, Y. Luo, et al. (2017) Deep attractor network for single-microphone speaker separation. In Proc. ICASSP 2017, pp. 246–250. Cited by: §1.
  • [8] A. Ephrat, I. Mosseri, et al. (2018) Looking to listen at the cocktail party: a speaker-independent audio-visual model for speech separation. arXiv preprint arXiv:1804.03619. Cited by: §1.
  • [9] C. Févotte, R. Gribonval, et al. (2005) BSS_EVAL toolbox user guide–revision 2.0. Cited by: §1.
  • [10] E. A. Habets and S. Gannot (2007) Generating sensor signals in isotropic noise fields. The Journal of the Acoustical Society of America 122 (6), pp. 3464–3470. Cited by: §4.1.
  • [11] J. R. Hershey, Z. Chen, and S. others (2016) Deep clustering: discriminative embeddings for segmentation and separation. In Proc. ICASSP 2016, pp. 31–35. Cited by: §1, §1.
  • [12] T. Hori, S. Araki, et al. (2012-02) Low-latency real-time meeting recognition and understanding using distant microphones and omni-directional camera. IEEE Transactions on Audio, Speech, and Language Processing 20 (2), pp. 499–513. External Links: Document, ISSN Cited by: §2.
  • [13] A. Janin, D. Baron, et al. (2003-04) The icsi meeting corpus. In Proc. ICASSP 2003, Vol. 1, pp. I–I. External Links: Document, ISSN Cited by: §3.2.
  • [14] N. Kanda, S. Horiguchi, et al. (2019-09) Simultaneous Speech Recognition and Speaker Diarization for Monaural Dialogue Recordings with Target-Speaker Acoustic Models. arXiv e-prints, pp. arXiv:1909.08103. External Links: 1909.08103 Cited by: §2.
  • [15] J. Le Roux, S. Wisdom, et al. (2019) SDR–half-baked or well done?. In Proc. ICASSP 2019, pp. 626–630. Cited by: §1.
  • [16] Z. Li, Y. Song, et al. (2019) Listening and grouping: an online autoregressive approach for monaural speech separation. IEEE/ACM Transactions on TASLP 27 (4), pp. 692–703. Cited by: §1.
  • [17] L. Lu, X. Xiao, et al. (2019) Pykaldi2: yet another speech toolkit based on kaldi and pytorch. arXiv preprint arXiv:1907.05955. Cited by: §3.3.1.
  • [18] Y. Luo and N. Mesgarani (2018) Tasnet: time-domain audio separation network for real-time, single-channel speech separation. In Proc. ICASSP 2018, Cited by: §1.
  • [19] Y. Luo, Z. Chen, et al. (2019) Dual-path rnn: efficient long sequence modeling for time-domain single-channel speech separation. External Links: 1910.06379 Cited by: §1.
  • [20] Y. Luo and N. Mesgarani (2019) Conv-tasnet: surpassing ideal time–frequency magnitude masking for speech separation. IEEE/ACM TASLP. Cited by: §1, §1.
  • [21] V. Panayotov, G. Chen, et al. (2015) Librispeech: an asr corpus based on public domain audio books. In Proc. ICASSP 2015, pp. 5206–5210. Cited by: §4.1.
  • [22] L. Perotin, R. Serizel, et al. (2018)

    Multichannel speech separation with recurrent neural networks from high-order ambisonics recordings

    In Proc. ICASSP 2018, Cited by: §1.
  • [23] D. Povey, A. Ghoshal, et al. (2011) The kaldi speech recognition toolkit. In Proc. ASRU, Cited by: §3.3.1.
  • [24] Z. Shi, H. Lin, et al. (2019) FurcaNeXt: end-to-end monaural speech separation with dynamic gated dilated temporal convolutional networks. arXiv preprint arXiv:1902.04891. Cited by: §1.
  • [25] T. v. Neumann, K. Kinoshita, et al. (2019-05) All-neural online source separation, counting, and diarization for meeting analysis. In Proc. ICASSP), Vol. , pp. 91–95. External Links: Document, ISSN Cited by: §2.
  • [26] M. Van Segbroeck, A. Zaid, et al. (2019) DiPCo–dinner party corpus. arXiv preprint arXiv:1909.13447. Cited by: §3.2.
  • [27] K. Wang, F. Soong, et al. (2019) A pitch-aware approach to single-channel speech separation. In Proc. ICASSP 2019, pp. 296–300. Cited by: §1.
  • [28] Q. Wang, H. Muckenhirn, et al. (2018) Voicefilter: targeted voice separation by speaker-conditioned spectrogram masking. arXiv preprint arXiv:1810.04826. Cited by: §1.
  • [29] Z. Wang, J. Le Roux, et al. (2018) Alternative objective functions for deep clustering. In Proc. ICASSP 2018, pp. 686–690. Cited by: §1.
  • [30] Z. Wang, J. Le Roux, et al. (2018) Multi-channel deep clustering: discriminative spectral and spatial embeddings for speaker-independent speech separation. Cited by: §1.
  • [31] G. Wichern, J. Antognini, et al. (2019) WHAM!: extending speech separation to noisy environments. arXiv preprint arXiv:1907.01160. Cited by: §3.2, §3.2.
  • [32] X. Xiao, Z. Chen, et al. (2019) Single-channel speech extraction using speaker inventory and attention network. In Proc. ICASSP 2019, pp. 86–90. Cited by: §1.
  • [33] C. Xu, W. Rao, et al. (2019) Optimization of speaker extraction neural network with magnitude and temporal spectrum approximation loss. In Proc. ICASSP 2019, pp. 6990–6994. Cited by: §1.
  • [34] T. Yoshioka, H. Erdogan, et al. (2018) Recognizing overlapped speech in meetings: a multichannel separation approach using neural networks. In Proc. Interspeech, pp. 3038–3042. Cited by: §4.1.
  • [35] T. Yoshioka, I. Abramovski, et al. (2019) Advances in online audio-visual meeting transcription. In Proc. ASRU, Cited by: §2, §3.2, §3.2, §4.1, §4.3.
  • [36] Yoshioka,Takuya, Erdogan,Hakan, et al. (2018) Multi-microphone neural speech separation for far-field multi-talker speech recognition. In Proc. ICASSP 2018, Cited by: §1, §1, §2, §4.2.
  • [37] D. Yu, M. Kolbæk, et al. (2017) Permutation invariant training of deep models for speaker-independent multi-talker speech separation. In Proc. ICASSP 2017, pp. 241–245. Cited by: §1.
  • [38] H. Zhao, C. Gan, A. Rouditchenko, et al. (2018) The sound of pixels. arXiv preprint arXiv:1804.03160. Cited by: §1.
  • [39] Y. Zhao, Z. Wang, et al. (2018) Two-stage deep learning for noisy-reverberant speech enhancement. IEEE/ACM TASLP 27 (1), pp. 53–62. Cited by: §1.
  • [40] K. Zmolikova, M. Delcroix, et al. (2017) Speaker-aware neural network based beamformer for speaker extraction in speech mixtures. In Interspeech, Cited by: §1.