pith. sign in

arxiv: 1906.10876 · v1 · pith:J3W5P7EAnew · submitted 2019-06-26 · 💻 cs.CL · cs.SD· eess.AS

Auxiliary Interference Speaker Loss for Target-Speaker Speech Recognition

Naoyuki Kanda , Shota Horiguchi , Ryoichi Takashima , Yusuke Fujita , Kenji Nagamatsu , Shinji Watanabe This is my paper

Pith reviewed 2026-05-25 16:09 UTC · model grok-4.3

classification 💻 cs.CL cs.SDeess.AS
keywords target-speaker ASRauxiliary lossspeaker separationmixed speechinterference speakersword error ratemonaural mixture
0
0 comments X

The pith

An auxiliary loss that also recognizes interference speakers improves target-speaker ASR from mixed audio.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper proposes an auxiliary loss for target-speaker automatic speech recognition that processes a monaural mixture of speakers given a short target-speaker sample. The loss adds a term that maximizes transcription accuracy on the non-target speakers during training. The goal is to produce internal representations that better separate the target speaker from the mixture. On two-speaker test mixtures the method lowers word error rate by 6.6 percent relative to a strong lattice-free maximum mutual information baseline. The same auxiliary output branch can also transcribe the interference speakers as a secondary task.

Core claim

The auxiliary interference speaker loss, by additionally maximizing ASR accuracy on interference speakers, regularizes the network to achieve better speaker separation representations and thereby higher accuracy on the target-speaker ASR task.

What carries the argument

Auxiliary interference speaker loss added to the training objective.

If this is right

  • Word error rate falls from 18.06 percent to 16.87 percent on the two-speaker test set.
  • The auxiliary output branch can be used directly as a secondary ASR system for interference speakers.
  • The improvement holds across a range of signal-to-interference ratio conditions.
  • The approach starts from a lattice-free maximum mutual information baseline that already outperforms a normal clean-speech ASR by a large margin.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same auxiliary-loss idea could be applied to mixtures containing more than two speakers by adding one branch per additional speaker.
  • The technique may reduce reliance on separate speaker-separation front-ends in fully end-to-end multi-speaker systems.
  • Gains could compound if the underlying single-speaker ASR model continues to improve.

Load-bearing premise

Training the model to recognize interference speakers will improve the internal separation of the target speaker rather than introduce conflicting gradients.

What would settle it

Train the model with the auxiliary loss and measure whether target-speaker word error rate on held-out mixtures fails to drop below the 18.06 percent baseline.

Figures

Figures reproduced from arXiv: 1906.10876 by Kenji Nagamatsu, Naoyuki Kanda, Ryoichi Takashima, Shinji Watanabe, Shota Horiguchi, Yusuke Fujita.

Figure 1
Figure 1. Figure 1: Overview of target-speaker AM architecture with aux￾iliary interference speaker loss. Auxiliary networks for inter￾ference speaker loss are only used in training and normally re￾moved in the decoding phase. A number with an arrow indicates a time splicing index, which forms the basis of a time-delay neu￾ral network (TDNN) [26]. The input features were advanced by five frames, which has the same effect as r… view at source ↗
Figure 2
Figure 2. Figure 2: Model architectures with early, middle, and late split￾ting. Note that the middle splitting model was used as the de￾fault model in other experiments [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Comparison among model architectures in [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
read the original abstract

In this paper, we propose a novel auxiliary loss function for target-speaker automatic speech recognition (ASR). Our method automatically extracts and transcribes target speaker's utterances from a monaural mixture of multiple speakers speech given a short sample of the target speaker. The proposed auxiliary loss function attempts to additionally maximize interference speaker ASR accuracy during training. This will regularize the network to achieve a better representation for speaker separation, thus achieving better accuracy on the target-speaker ASR. We evaluated our proposed method using two-speaker-mixed speech in various signal-to-interference-ratio conditions. We first built a strong target-speaker ASR baseline based on the state-of-the-art lattice-free maximum mutual information. This baseline achieved a word error rate (WER) of 18.06% on the test set while a normal ASR trained with clean data produced a completely corrupted result (WER of 84.71%). Then, our proposed loss further reduced the WER by 6.6% relative to this strong baseline, achieving a WER of 16.87%. In addition to the accuracy improvement, we also showed that the auxiliary output branch for the proposed loss can even be used for a secondary ASR for interference speakers' speech.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The paper proposes an auxiliary interference-speaker loss for target-speaker ASR on monaural mixtures. Given a short target-speaker enrollment, the method trains a network (based on lattice-free MMI) to transcribe the target while an auxiliary branch maximizes ASR accuracy on the interference speaker(s). The authors claim this auxiliary objective regularizes shared layers toward better speaker-separation representations, yielding a 6.6% relative WER reduction (18.06% → 16.87%) on two-speaker test mixtures; they also note that the auxiliary branch can serve as a secondary ASR for interference speech.

Significance. If the reported WER reduction is reproducible and the regularization mechanism is confirmed, the approach would offer a lightweight multi-task training recipe that improves target-speaker ASR without architectural changes. The use of a strong LF-MMI baseline and the secondary utility of the auxiliary branch are practical strengths; however, the absence of loss-weighting details, gradient diagnostics, or controls for capacity versus regularization effects limits the result's immediate impact on the field.

major comments (2)
  1. [Abstract / auxiliary loss description] Abstract and methods description of the auxiliary loss: the central claim that 'additionally maximizing interference speaker ASR accuracy during training ... will regularize the network to achieve a better representation for speaker separation' is presented without any reported loss weighting schedule, gradient-alignment measurements between the primary LF-MMI and auxiliary objectives, or ablation that isolates regularization from incidental multi-task or capacity effects. The 6.6% relative WER drop is therefore compatible with multiple alternative explanations.
  2. [Evaluation / results] Evaluation section: no auxiliary-task WER trajectory, correlation between auxiliary and target WER, or statistical significance tests on the 18.06% → 16.87% difference are provided, making it impossible to assess whether the observed improvement is robust or merely within run-to-run variance.
minor comments (1)
  1. [Abstract] The abstract states the baseline WER as 18.06% and the proposed result as 16.87% but does not specify the exact test-set size or number of runs, which would aid reproducibility.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on the auxiliary loss mechanism and evaluation details. We address each major comment below and will revise the manuscript to improve clarity and transparency.

read point-by-point responses

  1. Referee: [Abstract / auxiliary loss description] Abstract and methods description of the auxiliary loss: the central claim that 'additionally maximizing interference speaker ASR accuracy during training ... will regularize the network to achieve a better representation for speaker separation' is presented without any reported loss weighting schedule, gradient-alignment measurements between the primary LF-MMI and auxiliary objectives, or ablation that isolates regularization from incidental multi-task or capacity effects. The 6.6% relative WER drop is therefore compatible with multiple alternative explanations.

    Authors: We acknowledge that the original manuscript does not report the loss weighting schedule, gradient diagnostics, or ablations isolating regularization from capacity effects. The auxiliary loss was combined with the primary LF-MMI loss using a fixed weighting factor during training; we will specify this value and the exact training configuration in the revised methods section. We will also add a brief discussion noting alternative explanations for the WER improvement while retaining the claim that the auxiliary objective encourages better separation, supported by the demonstrated utility of the auxiliary branch for interference ASR. This will be a partial revision focused on clarification rather than new experiments. revision: partial

  2. Referee: [Evaluation / results] Evaluation section: no auxiliary-task WER trajectory, correlation between auxiliary and target WER, or statistical significance tests on the 18.06% → 16.87% difference are provided, making it impossible to assess whether the observed improvement is robust or merely within run-to-run variance.

    Authors: We agree these analyses are absent. We will add the auxiliary-task WER trajectory during training and the correlation with target-speaker WER if the per-epoch logs are recoverable. For the WER difference, we will include a statement that the 6.6% relative reduction was observed on a single test-set evaluation and that run-to-run variance was not quantified. We will also note consistency of gains across SIR conditions as supporting context. This constitutes a partial revision by adding available details and acknowledging limitations. revision: partial

Circularity Check

0 steps flagged

No circularity: empirical multi-task training result with no self-referential derivation

full rationale

The paper proposes an auxiliary loss for multi-task training and reports an empirical WER reduction (18.06% to 16.87%) on target-speaker ASR. No derivation chain exists that reduces a claimed prediction or first-principles result to its own inputs by construction; the regularization mechanism is stated as an intended effect of the loss rather than derived from equations that loop back to fitted parameters or self-citations. The result is an observed outcome of training, not a tautological renaming or fitted-input prediction. Self-citations, if present, are not load-bearing for any central claim.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the domain assumption that joint optimization of target and interference ASR objectives yields improved speaker separation representations. No free parameters or invented entities are mentioned in the abstract. Standard neural network training assumptions apply.

axioms (1)
  • domain assumption Jointly optimizing target-speaker and interference-speaker ASR objectives produces better internal representations for speaker separation than target-only training.
    This premise is invoked to justify why the auxiliary loss improves the main task.

pith-pipeline@v0.9.0 · 5768 in / 1283 out tokens · 47098 ms · 2026-05-25T16:09:55.929410+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

49 extracted references · 49 canonical work pages · 1 internal anchor

  1. [1]

    wsj0- 2mix

    Introduction Thanks to the recent advances in deep-learning [1–3], the ac cu- racy of automatic speech recognition (ASR) for some dataset s have become close to (e.g., Switchboard [4–6]) or beyond (e. g., LibriSpeech in [7] and [8]) the level of human transcribers. However, despite this progress, the accuracy of multi-talk er speech recognition is still v...

    work page

  2. [2]

    Auxiliary Interference Speaker Loss In this section, we explain our proposed method and its use with an LF-MMI-based AM due to its state-of-the-art perfor- mance [23, 27, 28]. However, it should be noted that our work can be extended to any kind of ASR loss like cross entropy (CE) [1], state-level minimum Bayes risk (sMBR) [29, 30], an d lattice-free sMBR...

    work page

  3. [3]

    dev-clean

    Evaluation 3.1. Experiment with LibriSpeech 3.1.1. Experimental settings For our primary evaluation, we used the LibriSpeech corpus [36], which consists of about 1,000 hours of read-aloud En- glish speech. In this study, we used 100 hours of the clean portion of the dataset for training the AMs. For evaluation, we used “dev-clean” and “test-clean” in acco...

    work page

  4. [4]

    Prepare a list of speech samples (main list), which is the main target of ASR

    work page

  5. [5]

    Shuffle the main list to create a second list under the con- straint that the same speaker does not appear in the same line in the main and second lists

    work page

  6. [6]

    For training data, we randomly sampled an SIR value from uniform distribution between -10 dB and 10 dB for each mixture

    Mix the audio in the main list and the second list one-by- one, with a specific SIR. For training data, we randomly sampled an SIR value from uniform distribution between -10 dB and 10 dB for each mixture. For the development and evaluation data, we generated data with an SIR of {10, 5, 0, -5, -10 } dB

    work page

  7. [7]

    Note that, in accordance with the protocol above, the speech of the target speaker could be much shorter or much longer than that of the interference speaker

    Only in the case of the training data, the volume of each mixed speech was randomly changed to enhance robust- ness for the volume difference. Note that, in accordance with the protocol above, the speech of the target speaker could be much shorter or much longer than that of the interference speaker. We intentionally sel ected this protocol because we bel...

    work page

  8. [8]

    clean AM

    with scales of 0.00005 and 0.1, respectively. The leaky hidden Markov model coefficient was set to 0.1. In addition, a backstitch technique [40] with a backstitch scale of 1.0 an d backstitch interval of 4 was used. For comparison purposes, we trained the AM without the proposed auxiliary loss, which corresponds to the original target-speaker model. We als...

    work page

  9. [9]

    We evaluated our proposed method using two-speaker-mixed speech in various SIR conditions

    Conclusions In this paper, we proposed a novel auxiliary loss function fo r target-speaker ASR, in which it attempts to maximize interf er- ence speaker ASR accuracy. We evaluated our proposed method using two-speaker-mixed speech in various SIR conditions. We first built a strong target-speaker ASR baseline based on the state-of-the-art LF-MMI, achieving ...

    work page

  10. [10]

    Conversational speech transc ription using context-dependent deep neural networks,

    F. Seide, G. Li, and D. Y u, “Conversational speech transc ription using context-dependent deep neural networks,” in Proc. INTER- SPEECH, 2011, pp. 437–440

    work page 2011

  11. [11]

    Context-depende nt pre-trained deep neural networks for large-vocabulary spe ech recognition,

    G. E. Dahl, D. Y u, L. Deng, and A. Acero, “Context-depende nt pre-trained deep neural networks for large-vocabulary spe ech recognition,” IEEE Trans. on ASLP , vol. 20, no. 1, pp. 30–42, 2012

    work page 2012

  12. [12]

    Deep neural networks for acoustic modeling in speech recognition: The shared views of four research groups,

    G. Hinton, L. Deng, D. Y u, G. E. Dahl, A.-r. Mohamed, N. Jai tly, A. Senior, V . V anhoucke, P . Nguyen, T. N. Sainath, and B. Kings- bury, “Deep neural networks for acoustic modeling in speech recognition: The shared views of four research groups,” Signal Processing Magazine, IEEE, vol. 29, no. 6, pp. 82–97, 2012

    work page 2012

  13. [13]

    Achieving Human Parity in Conversational Speech Recognition

    W. Xiong, J. Droppo, X. Huang, F. Seide, M. Seltzer, A. Sto lcke, D. Y u, and G. Zweig, “Achieving human parity in conversation al speech recognition,” arXiv preprint arXiv:1610.05256, 2016

    work page internal anchor Pith review Pith/arXiv arXiv 2016

  14. [14]

    English conversational telephone speech recognition by h umans and machines,

    G. Saon, G. Kurata, T. Sercu, K. Audhkhasi, S. Thomas, D. D im- itriadis, X. Cui, B. Ramabhadran, M. Picheny, L.-L. Lim et al. , “English conversational telephone speech recognition by h umans and machines,” Proc. INTERSPEECH, pp. 132–136, 2017

    work page 2017

  15. [15]

    Toward human parity in conversa- tional speech recognition,

    W. Xiong, J. Droppo, X. Huang, F. Seide, M. L. Seltzer, A. S tol- cke, D. Y u, and G. Zweig, “Toward human parity in conversa- tional speech recognition,” IEEE/ACM Trans. on ASLP , vol. 25, no. 12, pp. 2410–2423, 2017

    work page 2017

  16. [16]

    Deep speech 2: End-to-end speech recognition in English an d Mandarin,

    D. Amodei, S. Ananthanarayanan, R. Anubhai, J. Bai, E. Ba tten- berg, C. Case, J. Casper, B. Catanzaro, Q. Cheng, G. Chen et al., “Deep speech 2: End-to-end speech recognition in English an d Mandarin,” in Proc. ICML, 2016, pp. 173–182

    work page 2016

  17. [17]

    Lattice-free sta te-level minimum Bayes risk training of acoustic models,

    N. Kanda, Y . Fujita, and K. Nagamatsu, “Lattice-free sta te-level minimum Bayes risk training of acoustic models,” in Proc. IN- TERSPEECH, 2018, pp. 2923–2927

    work page 2018

  18. [18]

    Rec- ognizing overlapped speech in meetings: A multichannel sep ara- tion approach using neural networks,

    T. Y oshioka, H. Erdogan, Z. Chen, X. Xiao, and F. Alleva, “ Rec- ognizing overlapped speech in meetings: A multichannel sep ara- tion approach using neural networks,” in Proc. INTERSPEECH, 2018, pp. 3038–3042

    work page 2018

  19. [19]

    The fif th CHiME speech separation and recognition challenge: Datase t, task and baselines,

    J. Barker, S. Watanabe, E. Vincent, and J. Trmal, “The fif th CHiME speech separation and recognition challenge: Datase t, task and baselines,” in Proc. INTERSPEECH , 2018, pp. 1561– 1565

    work page 2018

  20. [20]

    The Hitachi/JHU CHiME-5 system: Advances in speech recognition for everyday home environments using mu lti- ple microphone arrays,

    N. Kanda, R. Ikeshita, S. Horiguchi, Y . Fujita, K. Nagam atsu, X. Wang, V . Manohar, N. E. Y . Soplin, M. Maciejewski, S.-J. Chen et al. , “The Hitachi/JHU CHiME-5 system: Advances in speech recognition for everyday home environments using mu lti- ple microphone arrays,” in Proc. CHiME-5, 2018, pp. 6–10

    work page 2018

  21. [21]

    Acoustic modeling for distant multi-talker s peech recognition with single- and multi-channel branches,

    N. Kanda, Y . Fujita, S. Horiguchi, R. Ikeshita, K. Nagam atsu, and S. Watanabe, “Acoustic modeling for distant multi-talker s peech recognition with single- and multi-channel branches,” in Proc. ICASSP, 2019

    work page 2019

  22. [22]

    Deep clus- tering: Discriminative embeddings for segmentation and se para- tion,

    J. R. Hershey, Z. Chen, J. Le Roux, and S. Watanabe, “Deep clus- tering: Discriminative embeddings for segmentation and se para- tion,” in Proc. ICASSP, 2016, pp. 31–35

    work page 2016

  23. [23]

    Deep attractor netwo rk for single-microphone speaker separation,

    Z. Chen, Y . Luo, and N. Mesgarani, “Deep attractor netwo rk for single-microphone speaker separation,” in Proc. ICASSP , 2017, pp. 246–250

    work page 2017

  24. [24]

    Recognizing multi-talker speech with permutation invariant training,

    D. Y u, X. Chang, and Y . Qian, “Recognizing multi-talker speech with permutation invariant training,” in Proc. INTERSPEECH , 2017, pp. 2456–2460

    work page 2017

  25. [25]

    Progressive joint modeling in unsupervised sing le- channel overlapped speech recognition,

    Z. Chen, J. Droppo, J. Li, W. Xiong, Z. Chen, J. Droppo, J. Li, and W. Xiong, “Progressive joint modeling in unsupervised sing le- channel overlapped speech recognition,” IEEE/ACM Trans. on ASLP, vol. 26, no. 1, pp. 184–196, 2018

    work page 2018

  26. [26]

    End-to-end multi-speaker speech recognition,

    S. Settle, J. Le Roux, T. Hori, S. Watanabe, and J. R. Hers hey, “End-to-end multi-speaker speech recognition,” in Proc. ICASSP, 2018, pp. 4819–4823

    work page 2018

  27. [27]

    A purely end-to-end system for multi-speaker speech recogni tion,

    H. Seki, T. Hori, S. Watanabe, J. Le Roux, and J. R. Hershe y, “A purely end-to-end system for multi-speaker speech recogni tion,” in Proc. ACL, 2018, pp. 2620–2630

    work page 2018

  28. [28]

    End-to-end monaural multi-speaker asr system without pretraining,

    X. Chang, Y . Qian, K. Y u, and S. Watanabe, “End-to-end monaural multi-speaker asr system without pretraining,” i n Proc. ICASSP, 2019

    work page 2019

  29. [29]

    Permutation invari- ant training of deep models for speaker-independent multi- talker speech separation,

    D. Y u, M. Kolbæk, Z.-H. Tan, and J. Jensen, “Permutation invari- ant training of deep models for speaker-independent multi- talker speech separation,” in Proc. ICASSP, 2017, pp. 241–245

    work page 2017

  30. [30]

    Speaker-aware neural network based beam- former for speaker extraction in speech mixtures,

    K. Zmolikova, M. Delcroix, K. Kinoshita, T. Higuchi, A. Ogawa, and T. Nakatani, “Speaker-aware neural network based beam- former for speaker extraction in speech mixtures,” in Proc. IN- TERSPEECH, 2017

    work page 2017

  31. [31]

    Single channel target speaker extraction and recog- nition with speaker beam,

    M. Delcroix, K. Zmolikova, K. Kinoshita, A. Ogawa, and T. Nakatani, “Single channel target speaker extraction and recog- nition with speaker beam,” in ICASSP, 2018, pp. 5554–5558

    work page 2018

  32. [32]

    Purely sequence-trained neu- ral networks for ASR based on lattice-free MMI,

    D. Povey, V . Peddinti, D. Galvez, P . Ghahrmani, V . Manoh ar, X. Na, Y . Wang, and S. Khudanpur, “Purely sequence-trained neu- ral networks for ASR based on lattice-free MMI,” inProc. INTER- SPEECH, 2016, pp. 2751–2755

    work page 2016

  33. [33]

    Single-channel multi-speaker separation using deep clus tering,

    Y . Isik, J. Le Roux, Z. Chen, S. Watanabe, and J. R. Hershe y, “Single-channel multi-speaker separation using deep clus tering,” in Proc. INTERSPEECH, 2016, pp. 545–549

    work page 2016

  34. [34]

    Single-channel multi-tal ker speech recognition with permutation invariant training,

    Y . Qian, X. Chang, and D. Y u, “Single-channel multi-tal ker speech recognition with permutation invariant training,” Speech Communication, vol. 104, pp. 1–11, 2018

    work page 2018

  35. [35]

    A time delay ne ural network architecture for efficient modeling of long tempora l con- texts,

    V . Peddinti, D. Povey, and S. Khudanpur, “A time delay ne ural network architecture for efficient modeling of long tempora l con- texts,” in Proc. INTERSPEECH, 2015, pp. 3214–3218

    work page 2015

  36. [36]

    Investigation o f lattice- free maximum mutual information-based acoustic models wit h sequence-level Kullback-Leibler divergence,

    N. Kanda, Y . Fujita, and K. Nagamatsu, “Investigation o f lattice- free maximum mutual information-based acoustic models wit h sequence-level Kullback-Leibler divergence,” in Proc. ASRU , 2017, pp. 69–76

    work page 2017

  37. [37]

    Sequence distil lation for purely sequence trained acoustic models,

    N. Kanda, Y . Fujita, and K. Nagamatsu, “Sequence distil lation for purely sequence trained acoustic models,” in Proc. ICASSP, 2018, pp. 5964–5968

    work page 2018

  38. [38]

    Sequen ce- discriminative training of deep neural networks,

    K. V esel` y, A. Ghoshal, L. Burget, and D. Povey, “Sequen ce- discriminative training of deep neural networks,” in Proc. INTER- SPEECH, 2013, pp. 2345–2349

    work page 2013

  39. [39]

    Error back propagation for se- quence training of context-dependent deep networks for con ver- sational speech transcription,

    H. Su, G. Li, D. Y u, and F. Seide, “Error back propagation for se- quence training of context-dependent deep networks for con ver- sational speech transcription,” in Proc. ICASSP, 2013, pp. 6664– 6668

    work page 2013

  40. [40]

    Con- nectionist temporal classification: labelling unsegmente d se- quence data with recurrent neural networks,

    A. Graves, S. Fern´ andez, F. Gomez, and J. Schmidhuber, “Con- nectionist temporal classification: labelling unsegmente d se- quence data with recurrent neural networks,” in Proc. ICML , 2006, pp. 369–376

    work page 2006

  41. [41]

    End-t o- end continuous speech recognition using attention-based recurrent NN: First results,

    J. Chorowski, D. Bahdanau, K. Cho, and Y . Bengio, “End-t o- end continuous speech recognition using attention-based recurrent NN: First results,” Proc. NIPS workshop on Deep Learning and Representation Learning, 2014

    work page 2014

  42. [42]

    Listen, atten d and spell: A neural network for large vocabulary conversati onal speech recognition,

    W. Chan, N. Jaitly, Q. Le, and O. Vinyals, “Listen, atten d and spell: A neural network for large vocabulary conversati onal speech recognition,” in Proc. ICASSP, 2016, pp. 4960–4964

    work page 2016

  43. [43]

    Front-end factor analysis for speaker verification,

    N. Dehak, P . J. Kenny, R. Dehak, P . Dumouchel, and P . Ouel let, “Front-end factor analysis for speaker verification,” IEEE Trans. on ASLP, vol. 19, no. 4, pp. 788–798, 2011

    work page 2011

  44. [44]

    Speaker adap- tation of neural network acoustic models using i-vectors,

    G. Saon, H. Soltau, D. Nahamoo, and M. Picheny, “Speaker adap- tation of neural network acoustic models using i-vectors,” in Proc. ASRU, 2013, pp. 55–59

    work page 2013

  45. [45]

    Lib- rispeech: an ASR corpus based on public domain audio books,

    V . Panayotov, G. Chen, D. Povey, and S. Khudanpur, “Lib- rispeech: an ASR corpus based on public domain audio books,” in Proc. ICASSP, 2015

    work page 2015

  46. [46]

    The Kaldi speech recognition toolkit,

    D. Povey, A. Ghoshal, G. Boulianne, L. Burget, O. Glembe k, N. Goel, M. Hannemann, P . Motl´ ıˇ cek, Y . Qian, P . Schwarzet al., “The Kaldi speech recognition toolkit,” in Proc. ASRU, 2011

    work page 2011

  47. [47]

    Phoneme recognition using time-delay neural networks,

    A. Waibel, T. Hanazawa, G. Hinton, K. Shikano, and K. J. L ang, “Phoneme recognition using time-delay neural networks,” IEEE Trans. ASSP, vol. 37, no. 3, pp. 328–339, 1989

    work page 1989

  48. [48]

    Long short-term mem ory,

    S. Hochreiter and J. Schmidhuber, “Long short-term mem ory,” Neural computation, vol. 9, no. 8, pp. 1735–1780, 1997

    work page 1997

  49. [49]

    Backstitch: Counteracting finite-sample bias via neg ative steps,

    Y . Wang, V . Peddinti, H. Xu, X. Zhang, D. Povey, and S. Khudan- pur, “Backstitch: Counteracting finite-sample bias via neg ative steps,” in Proc. INTERSPEECH, 2017, pp. 1631–1635

    work page 2017