Pith. sign in

REVIEW 4 major objections 6 minor 31 references

DurIAN: Duration Informed Attention Network For Multimodal Synthesis

T0 review · 4 major / 6 minor · reviewed 2026-08-14 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read DurIAN matches Tacotron 2's speech naturalness while eliminating skip-and-repeat errors.

desk verdict A solid industrial TTS paper with one directly measured headline result (MOS on par with Tacotron 2, 0% skip/repeat), but robustness evidence covers only skip/repeat and the multimodal/style claims are demo-based. read the letter →

arxiv 1909.01700 v2 pith:NNWO3T2Q submitted 2019-09-04 cs.CL cs.CVcs.SDeess.AS

Chengzhu Yu , Heng Lu , Na Hu , Meng Yu , Chao Weng , Kun Xu , Peng Liu , Deyi Tuo
show 4 more authors
Shiyin Kang Guangzhi Lei Dan Su Dong Yu
This is my paper · ORCID
classification cs.CLcs.CVcs.SDeess.AS
keywords text-to-speechdurationinformedattentionrobustspeechsynthesisalignmentmodelmultimodalWaveRNNstylecontrolforced
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

DurIAN sets out to show that the main source of instability in modern text-to-speech—words that get skipped or repeated at random—comes from the attention mechanism, not from the rest of the neural pipeline. To prove this, the paper replaces end-to-end attention with an explicit alignment model driven by predicted phoneme durations, keeping everything else from the Tacotron 2 recipe. In listening tests on male and female Mandarin voices, the resulting speech is statistically indistinguishable in naturalness from Tacotron 2 (MOS 4.11 vs 4.10 male; 4.26 vs 4.28 female), while word skip/repeat errors drop from 2% to 0% on 1,000 unseen utterances. If right, this means robust, production-safe neural TTS can keep the naturalness gains of end-to-end systems without their characteristic failure mode. The same duration model also synchronizes speech with facial expression, and a multi-band WaveRNN makes the vocoder run faster than real time on a single CPU core.

What carries the argument

The central mechanism is the alignment model: a duration predictor plus a state-expansion operation that replicates each phoneme's encoder hidden state for as many acoustic frames as that phoneme lasts, with the in-phoneme position normalized to a value between 0 and 1 and appended to the state. This replaces the attention context that end-to-end systems use, and it is what prevents the decoder from jumping backward or forward across the input. A skip encoder encodes phonemes together with Chinese prosodic boundary symbols but drops the boundary states before expansion; a separate duration model predicts phoneme durations at inference, optionally conditioned on a scaled style embedding for fine-grained expressiveness control.

What would settle it

A concrete check: on an out-of-domain test set, synthesize with DurIAN, force-align the output audio to the text, and count both (a) words skipped or repeated and (b) phoneme boundaries whose timing deviates from a human-aligned reference by more than about 40 ms. If (b) is non-negligible while (a) is zero, the paper's robustness claim is incomplete—DurIAN would have traded one alignment failure mode for another.

Watch

Extended reading notes

Core claim

The paper identifies the end-to-end attention mechanism as the cause of the unpredictable skipped- or repeated-word artifacts seen in systems like Tacotron 2. It introduces DurIAN, an autoregressive model whose encoder states are expanded according to phoneme durations instead of attention: during training, durations come from forced alignment of the input phonemes to the acoustic frames, and during synthesis they are predicted by a separately trained bidirectional-LSTM duration model. The decoder then consumes frame-aligned encoder states plus a relative position in each phoneme, and a post-net refines the predicted mel-spectrogram. In head-to-head tests, DurIAN's naturalness matches Tacotron 2's within statistical noise while producing no skip/repeat errors on 1,000 unseen utterances, against 2% for Tacotron 2 and 4% for Deep Voice 3. The paper also shows the same duration-model interface can drive facial-expression features synchronized with speech even when voice and face come from different data sources, and that a four-band WaveRNN vocoder with 8-bit quantization cuts inference cost from about 9.8 to 3.6 GFLOPS and runs about six times faster than real time on one CPU core.

Load-bearing premise

The load-bearing premise is that the forced-aligner-derived phoneme durations used as training targets are accurate, and that the trained duration model keeps text and audio aligned on sentences outside its training distribution; if either fails, DurIAN would replace skipped or repeated words with mis-timed or mis-aligned speech, a failure mode the reported skip/repeat metric does not measure.

Editorial extensions

If this is right

  • Production systems can get end-to-end-grade naturalness on out-of-domain text, because the decoder is forced to follow a monotonic duration schedule and cannot jump over or revisit input positions.
  • The naturalness advantage of neural TTS does not depend on end-to-end attention itself: DurIAN's on-par MOS shows the encoder, autoregressive decoder, post-net, and neural vocoder carry the quality.
  • Voice and face can be paired freely: because a shared duration model drives both modalities, a speech synthesis model and a facial-expression model trained on different datasets can be synchronized without parallel speech–face recordings.
  • Discrete style labels are enough for continuous control: scaling a learned style embedding changes the intensity of a style at inference time, so a model trained on few labeled categories can interpolate between them.
  • Neural vocoding can run in real time on commodity CPUs: the 4-band WaveRNN with 8-bit quantization and SIMD reaches a real-time factor of 0.171 on one core, with no statistically significant quality drop.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A direct test the paper does not run is timing accuracy: force-align DurIAN's synthesized audio and measure whether predicted phoneme boundaries drift on long or out-of-domain sentences; a skip/repeat test would not catch a systematic rhythm error introduced by a biased duration model.
  • Because duration is an explicit variable, DurIAN invites prosody editing as a post-hoc control—users could stretch or compress individual phoneme durations at inference to change tempo or emphasis, a capability attention-based models make awkward.
  • The same alignment interface could drive other time-locked modalities—lip motion, gestures, or avatar body movement—whenever events can be attached to phoneme boundaries, generalizing the paper's speech-plus-face result.
  • A robustness extension could gate the duration model: on inputs where predicted durations have low confidence or high variance, fall back to attention or insert a correction step, addressing the out-of-distribution risk that the paper's held-out set may not exercise.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The paper proposes DurIAN, an autoregressive text-to-speech (TTS) framework that replaces the end-to-end attention mechanism of systems like Tacotron 2 with an alignment model based on phoneme durations. DurIAN combines a skip encoder for phoneme and prosodic-structure encoding, a separately trained duration predictor, an autoregressive decoder with post-net, a style-code mechanism for fine-grained style control, and a Multi-band WaveRNN vocoder for faster inference. The authors report that DurIAN matches Tacotron 2 in naturalness (MOS 4.11 vs. 4.10 for male speech; 4.26 vs. 4.28 for female speech), reduces word skip/repeat errors to 0% on 1000 utterances versus 2% for their Tacotron 2 baseline, speeds up WaveRNN inference from 1.337 to 0.171 RTF, and demo synchronized speech with facial expression. The paper also describes a duration-based method for synchronizing speech and facial expression without paired data.

Significance. If the reported results hold, the paper makes a useful and timely industrial point: an alignment model based on predicted durations can recover most of the naturalness of end-to-end TTS while avoiding the characteristic attention artifacts. The direct comparison against Tacotron 2 and a parametric BLSTM baseline is appropriate, and the 1000-utterance robustness test is a valuable practical stress test. The complexity analysis for Multi-band WaveRNN is explicit and checkable, and the multi-band shared-RNN idea is a plausible efficiency contribution. However, the paper's breadth exceeds its evaluation depth: the style-control and multimodal-synchronization claims are supported only by demo examples, and the core robustness claim is measured only for one artifact class. The manuscript is a solid systems paper, but several load-bearing claims need additional evidence before journal-level acceptance.

major comments (4)
  1. [Sec. 5.1, Table 2] The central robustness claim is supported only for word skipping/repeating errors. Because DurIAN replaces attention with a separately trained duration model (Sec. 2.2, Eq. (2)), its out-of-domain failure mode would be mistimed phoneme boundaries or shifted prosodic-boundary positions rather than skipped/repeated words. The 1000-utterance test reports only the occurrence rate of skip/repeat errors, so a system with 0% skip/repeat could still be badly misaligned; the paper should add direct alignment-quality evidence (e.g., duration-prediction error, boundary offsets, or a listening test for timing and rhythm) and preferably test on the same out-of-domain sentences used for the female MOS evaluation.
  2. [Sec. 5.1, Table 1] The claim that DurIAN and Tacotron 2 are 'on par' is not statistically supported: the text asserts 'No statistically significant difference can be observed' but reports no confidence intervals, significance test, or listener-level variance. With 40 (male) and 20 (out-of-domain female) test sentences, the MOS differences (4.11 vs. 4.10; 4.26 vs. 4.28) could easily be within rating noise. Please report per-sentence score distributions, bootstrapped confidence intervals, and a paired significance test before drawing the parity conclusion.
  3. [Sec. 5.3 and 5.4] The fine-grained style control and multimodal synchronization results are presented only as demonstrations with demo URLs; no quantitative evaluation is reported. Section 3 explicitly relies on the 'presumption' that scaling a learned style embedding changes style intensity while preserving style identity, but no perceptual test validates this. Likewise, Sec. 5.4 gives no metric for audio-visual synchronization (e.g., timing offsets in frames). Since items 2 and 3 of the contribution list in Sec. 1 include these capabilities, a small controlled experiment (e.g., style-scale ratings, identity/speaker similarity, or sync-error measurement) is needed.
  4. [Sec. 2.2 and 5.1] The accuracy of the forced-alignment durations and of the duration predictor is never reported. Because duration prediction is the load-bearing component that replaces attention, the reader needs at least a few implementation details and validation numbers: the aligner type, the duration-model input features, and duration prediction error on the test sets (e.g., RMSE or frame-level error). Without this, it is difficult to assess whether DurIAN's alignment is trustworthy on unseen or out-of-domain text, which is exactly where the paper claims robustness.
minor comments (6)
  1. [Abstract vs. Sec. 5.2.1] The abstract says the Multi-band WaveRNN reduces complexity 'from 9.8 to 5.5 GFLOPS', but Sec. 5.2.1 computes 3.6 GFLOPS for NB=4; the abstract should be corrected.
  2. [Table 4] Table 4 lists two rows labeled '4-band WaveRNN (int8)' with scores 4.58 and 4.56; one of these is presumably the 4-band float model and should be labeled accordingly.
  3. [Table 2] Table 2 contains two 'Tacotron 2' rows with different skip/repeat rates (1% and 2%); please clarify which row is the system reproduced in this paper and which is taken from the original reference.
  4. [Eq. (5)] Equation (5) omits the frame index on the loss terms; it should read Σ_n |y_n − y'_n| + Σ_n |y_n − (y'_n + r_n)|.
  5. [Sec. 4.1] The 'quantitative loss learning mechanism' mentioned in Sec. 4.1 is not described or cited; please specify the method or provide a reference.
  6. [Appendix A and Sec. 4.2/5.2.2] There are several typos: 'Pseudo Quadratue' should be 'Pseudo Quadrature', 'alogorithm' should be 'algorithm', and 'If fact' should be 'In fact'.

Circularity Check

1 steps flagged · score 1.0 of 10

Robustness comparison is partly definitional because DurIAN removes attention by design; naturalness and complexity results are externally measured and self-contained.

  1. self definitional [Sec. 2.2 (Alignment Model) and Sec. 5.1 (Robustness evaluation, Table 2)]
    "In DurIAN, we replace the attention mechanism with an alignment model [15, 16], in which the alignment between the phoneme sequence and the target acoustic sequence is inferred from a phoneme duration prediction model. ... In this evaluation, we mainly focused on the word skipping and repeating errors commonly occur in the Tacotron 2 systems. ... These results clearly indicate that DurIAN is much more robust than Tacotron-2 and generated no error in this category."

    By Eq. (2), e_{1:T} = state_expand(h_{1:N'}, d_{1:N'}), encoder states are replicated deterministically from predicted durations; there is no attention distribution that can jump backward or forward and cause the standard attention-mediated word skip/repeat artifact named in Sec. 2.2. Thus the 0% skip/repeat rate in Table 2 is guaranteed by architecture choice rather than being a separate empirical prediction. The robustness test measures only this artifact class, not the mis-timed or mis-aligned boundaries that a biased duration model could produce, so the 'much more robust' conclusion is only established for the failure mode the design eliminates by construction.

full rationale

Most of the paper's claims are self-contained and benchmarked against external systems. The naturalness MOS results (Table 1) compare DurIAN with Tacotron 2 and a parametric baseline on held-out sentences; no fitted parameter is renamed as a prediction. The duration model is trained on forced-alignment targets and its quality is assessed through synthesized speech on unseen text, which is a genuine external test. The 9.8-to-3.6 GFLOPS reduction is an arithmetic consequence of Eq. (6) with stated architecture sizes, not a fitted or circular result. The style-control section explicitly frames its mechanism as a 'presumption' and demonstrates it with samples rather than deriving it from first principles. Citations to [15,16] for alignment models and [6,7] for Tacotron components are external technical references; the self-citation to the patent [13] is used only to establish priority and is not load-bearing for any scientific claim. The only near-circular element is the robustness comparison, which evaluates only word skip/repeat errors, the artifact class that removing attention eliminates by construction. Because that metric is partly definitional while the naturalness and efficiency claims are externally validated, the overall circularity is minimal.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The central claims rest on four domain assumptions that are stated but not independently verified: forced-alignment duration accuracy, the style-scaling presumption, the prosody-token benefit, and the transferability of speech durations to facial expressions. The only hand-set quantities are the style control scale and the subband count. No new physical or latent entity is postulated beyond the style code, which is a scaled embedding rather than a new entity.

free parameters (2)
  • Style control scale w = 1.0 during training; continuous at inference
    The method scales a learned style embedding by this user-set multiplier; the paper asserts magnitude corresponds to style intensity without calibrating the scale or validating it in listening tests (Sec. 3).
  • Number of subbands NB = 4 (compared to fullband NB=1)
    The reported GFLOPS reduction and RTF 0.171 are for this hand-chosen operating point; the complexity formula makes the recurrent cost scale as SR/NB, so the speed claim depends on this choice (Sec. 5.2.1).
assumptions (4)
  • domain assumption Forced alignment durations are accurate enough to serve as ground truth for state expansion and duration-model training.
    Sec. 2.2 states durations are 'obtained through forced alignment' and used to expand encoder states; alignment errors would propagate into the decoder training targets and the duration model.
  • ad hoc to paper Scaling a learned style embedding changes style intensity while preserving style identity.
    Sec. 3 explicitly calls this a 'presumption'; no derivation or experimental validation is provided.
  • domain assumption Prosodic boundary tokens (#1, #2, #3) improve out-of-domain generalization for Chinese synthesis.
    Sec. 2.1 asserts early experiments favored the skip encoder over linguistic features but reports no quantitative comparison.
  • domain assumption A speech-trained duration model can drive synchronized facial expressions without parallel speech and face data.
    Sec. 2.4 relies on this equivalence for duration-based synchronization; no synchronization accuracy or naturalness measure is reported.

how reviews work

0 comments
Cite this review

Pith. "Pith review of DurIAN: Duration Informed Attention Network For Multimodal Synthesis." pith.science (2026). https://pith.science/paper/NNWO3T2Q

@misc{pith2026190901700,
  author       = {Pith},
  title        = {Pith review of: DurIAN: Duration Informed Attention Network For Multimodal Synthesis},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NNWO3T2Q}},
  note         = {Machine review of arXiv:1909.01700}
}
read the original abstract

In this paper, we present a generic and robust multimodal synthesis system that produces highly natural speech and facial expression simultaneously. The key component of this system is the Duration Informed Attention Network (DurIAN), an autoregressive model in which the alignments between the input text and the output acoustic features are inferred from a duration model. This is different from the end-to-end attention mechanism used, and accounts for various unavoidable artifacts, in existing end-to-end speech synthesis systems such as Tacotron. Furthermore, DurIAN can be used to generate high quality facial expression which can be synchronized with generated speech with/without parallel speech and face data. To improve the efficiency of speech generation, we also propose a multi-band parallel generation strategy on top of the WaveRNN model. The proposed Multi-band WaveRNN effectively reduces the total computational complexity from 9.8 to 5.5 GFLOPS, and is able to generate audio that is 6 times faster than real time on a single CPU core. We show that DurIAN could generate highly natural speech that is on par with current state of the art end-to-end systems, while at the same time avoid word skipping/repeating errors in those systems. Finally, a simple yet effective approach for fine-grained control of expressiveness of speech and facial expression is introduced.

Figures

Figures reproduced from arXiv: 1909.01700 by the authors.

Figure 1
Figure 1. Model architecture of DurIAN. The model takes a sequence of symbols, including phonemes and prosodic [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. An illustration of how prosodic boundaries are inserted between input phonemes. The symbol [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Generation of the style code from style embedding. The style embedding is learned jointly with other [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Model architecture of Multi-band WaveRNN. [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Frequency response of 4-band Pseudo-QMF filter banks. [PITH_FULL_IMAGE:figures/full_fig_p011_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

31 extracted references · 18 canonical work pages

  1. [1]

    Unit selection in a concatenative speech synthesis system using a large speech database,

    A. J. Hunt and A. W. Black, “Unit selection in a concatenative speech synthesis system using a large speech database,” in 1996 IEEE International Conference on Acoustics, Speech, and Signal Processing Conference Proceedings, vol. 1, pp. 373–376, IEEE, 1996

    work page 1996

  2. [2]

    Automatically clustering similar units for unit selection in speech synthesis.,

    A. W. Black and P. A. Taylor, “Automatically clustering similar units for unit selection in speech synthesis.,” 1997

    work page 1997

  3. [3]

    Speech parameter generation algorithms for hmm-based speech synthesis,

    K. Tokuda, T. Yoshimura, T. Masuko, T. Kobayashi, and T. Kitamura, “Speech parameter generation algorithms for hmm-based speech synthesis,” in 2000 IEEE International Conference on Acoustics, Speech, and Signal Processing. Proceedings (Cat. No. 00CH37100), vol. 3, pp. 1315–1318, IEEE, 2000

    work page 2000

  4. [4]

    Statistical parametric speech synthesis,

    H. Zen, K. Tokuda, and A. W. Black, “Statistical parametric speech synthesis,”speech communication, vol. 51, no. 11, pp. 1039–1064, 2009

    2009

  5. [5]

    Statistical parametric speech synthesis using deep neural networks,

    H. Zen, A. Senior, and M. Schuster, “Statistical parametric speech synthesis using deep neural networks,” in2013 ieee international conference on acoustics, speech and signal processing , pp. 7962–7966, IEEE, 2013

    work page 2013

  6. [6]

    Tacotron: Towards end-to-end speech synthesis,

    Y . Wang, R. Skerry-Ryan, D. Stanton, Y . Wu, R. J. Weiss, N. Jaitly, Z. Yang, Y . Xiao, Z. Chen, S. Bengio,et al., “Tacotron: Towards end-to-end speech synthesis,”arXiv preprint arXiv:1703.10135, 2017

    arXiv 2017

  7. [7]

    Natural tts synthesis by conditioning wavenet on mel spectrogram predictions,

    J. Shen, R. Pang, R. J. Weiss, M. Schuster, N. Jaitly, Z. Yang, Z. Chen, Y . Zhang, Y . Wang, R. Skerrv-Ryan, et al., “Natural tts synthesis by conditioning wavenet on mel spectrogram predictions,” in2018 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , pp. 4779–4783, IEEE, 2018

    work page 2018

  8. [8]

    Close to human quality tts with transformer,

    N. Li, S. Liu, Y . Liu, S. Zhao, M. Liu, and M. Zhou, “Close to human quality tts with transformer,”arXiv preprint arXiv:1809.08895, 2018

    arXiv 2018

Show all 31 references
  1. [9]

    Clarinet: Parallel wave generation in end-to-end text-to-speech,

    W. Ping, K. Peng, and J. Chen, “Clarinet: Parallel wave generation in end-to-end text-to-speech,”arXiv preprint arXiv:1807.07281, 2018

    2018 arXiv

  2. [10]

    Char2wav: End-to-end speech synthesis,

    J. Sotelo, S. Mehri, K. Kumar, J. F. Santos, K. Kastner, A. Courville, and Y . Bengio, “Char2wav: End-to-end speech synthesis,” 2017

    2017

  3. [11]

    Deep voice 3: Scaling text-to-speech with convolutional sequence learning,

    W. Ping, K. Peng, A. Gibiansky, S. O. Arik, A. Kannan, S. Narang, J. Raiman, and J. Miller, “Deep voice 3: Scaling text-to-speech with convolutional sequence learning,”arXiv preprint arXiv:1710.07654, 2017

    2017 arXiv

  4. [12]

    Fastspeech: Fast, robust and controllable text to speech,

    Y . Ren, Y . Ruan, X. Tan, T. Qin, S. Zhao, Z. Zhao, and T.-Y . Liu, “Fastspeech: Fast, robust and controllable text to speech,” arXiv preprint arXiv:1905.09263, 2019

    1905 arXiv

  5. [13]

    Duration Informed Attention Network For Text-To-Speech Analysis,

    C. Yu, H. Lu, and D. Yu, “Duration Informed Attention Network For Text-To-Speech Analysis,”U.S. Provisional application, Pending, April, 2019

    2019

  6. [14]

    Efficient neural audio synthesis,

    N. Kalchbrenner, E. Elsen, K. Simonyan, S. Noury, N. Casagrande, E. Lockhart, F. Stimberg, A. v. d. Oord, S. Dieleman, and K. Kavukcuoglu, “Efficient neural audio synthesis,”arXiv preprint arXiv:1802.08435, 2018. 9 DurIAN: Duration Informed Attention Network For Multimodal Synthesis

    2018 arXiv

  7. [15]

    TTS Synthesis with Bidirectional LSTM based Recurrent Neural Networks,

    Y . Fan, Y . Qian, F. Xie, and F. K. Soong, “TTS Synthesis with Bidirectional LSTM based Recurrent Neural Networks,” INTERSPEECH, pp. 1964–1968, Singapore, September, 2014

    1964

  8. [16]

    Fast, compact, and high quality lstm-rnn based statistical parametric speech synthesizers for mobile devices,

    H. Zen, Y . Agiomyrgiannakis, N. Egberts, F. Henderson, and P. Szczepaniak, “Fast, compact, and high quality lstm-rnn based statistical parametric speech synthesizers for mobile devices,”arXiv preprint arXiv:1606.06061, 2016

    2016 arXiv

  9. [17]

    Grammar as a foreign language,

    O. Vinyals, Ł. Kaiser, T. Koo, S. Petrov, I. Sutskever, and G. Hinton, “Grammar as a foreign language,” in Advances in neural information processing systems , pp. 2773–2781, 2015

    2015

  10. [18]

    Style tokens: Unsupervised style modeling, control and transfer in end-to-end speech synthesis,

    Y . Wang, D. Stanton, Y . Zhang, R. Skerry-Ryan, E. Battenberg, J. Shor, Y . Xiao, F. Ren, Y . Jia, and R. A. Saurous, “Style tokens: Unsupervised style modeling, control and transfer in end-to-end speech synthesis,” arXiv preprint arXiv:1803.09017, 2018

    2018 arXiv

  11. [19]

    Hierarchical generative modeling for controllable speech synthesis,

    W.-N. Hsu, Y . Zhang, R. J. Weiss, H. Zen, Y . Wu, Y . Wang, Y . Cao, Y . Jia, Z. Chen, J. Shen,et al., “Hierarchical generative modeling for controllable speech synthesis,”arXiv preprint arXiv:1810.07217, 2018

    2018 arXiv

  12. [20]

    Disentangling correlated speaker and noise for speech synthesis via data augmentation and adversarial factorization,

    W.-N. Hsu, Y . Zhang, R. J. Weiss, Y .-A. Chung, Y . Wang, Y . Wu, and J. Glass, “Disentangling correlated speaker and noise for speech synthesis via data augmentation and adversarial factorization,” in ICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Si...

    2019

  13. [21]

    Wavenet: A generative model for raw audio,

    A. v. d. Oord, S. Dieleman, H. Zen, K. Simonyan, O. Vinyals, A. Graves, N. Kalchbrenner, A. Senior, and K. Kavukcuoglu, “Wavenet: A generative model for raw audio,”arXiv preprint arXiv:1609.03499, 2016

    2016 arXiv

  14. [22]

    Parallel wavenet: Fast high-fidelity speech synthesis,

    A. v. d. Oord, Y . Li, I. Babuschkin, K. Simonyan, O. Vinyals, K. Kavukcuoglu, G. v. d. Driessche, E. Lockhart, L. C. Cobo, F. Stimberg, et al., “Parallel wavenet: Fast high-fidelity speech synthesis,”arXiv preprint arXiv:1711.10433, 2017

    2017 arXiv

  15. [23]

    Lpcnet: Improving neural speech synthesis through linear prediction,

    J.-M. Valin and J. Skoglund, “Lpcnet: Improving neural speech synthesis through linear prediction,” inICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , pp. 5891–5895, IEEE, 2019

    2019

  16. [24]

    Improving FFTNet vocoder with noise shaping and subband approaches,

    T. Okamoto, T. Toda, Y . Shiga, and H. Kawai, “Improving FFTNet vocoder with noise shaping and subband approaches,” in 2018 IEEE Spoken Language Technology Workshop (SLT), pp. 304–311, IEEE, 2018

    2018

  17. [25]

    An investigation of subband WaveNet vocoder covering entire audible frequency range with limited acoustic features,

    T. Okamoto, K. Tachibana, T. Toda, Y . Shiga, and H. Kawai, “An investigation of subband WaveNet vocoder covering entire audible frequency range with limited acoustic features,” in2018 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , pp. 5654...

    2018

  18. [26]

    Facewarehouse: a 3D facial expression database for visual computing,

    C. Chen, Y . Weng, S. Zhou, Y . Tong, and K. Zhou, “Facewarehouse: a 3D facial expression database for visual computing,” IEEE Transactions on Visualization and Computer Graphics , pp. 413–425, 2014

    2014

  19. [27]

    Deep video portraits,

    H. Kim, P. Garrido, A. Tewari, W. Xu, J. Thies, M. Nießner, P. Perez, C. Richardt, M. Zollhöfer, and C. Theobalt, “Deep video portraits,”Siggraph, 2018

    2018

  20. [28]

    Text-based editing of talking-head video,

    O. Fried, A. Tewari, M. Zollhöfer, A. Finkelstein, E. Shechtman, D. B. Goldman, K. Genova, Z. Jin, C. Theobalt, and M. Agrawala, “Text-based editing of talking-head video,”ACM Transactions on Graphics, vol. V olume 38 Issue 4, July, 2019

    2019

  21. [29]

    Near-perfect-reconstruction pseudo-qmf banks,

    T. Nguyen, “Near-perfect-reconstruction pseudo-qmf banks,”IEEE Transactions on Signal Processing, V ol. 42, No.1,, 1994

    1994

  22. [30]

    An investigation of subband wavenet vocoder covering entire audible frequency range with limited acoustic features,

    T. Okamoto, K. Tachibana, T. Toda, Y . Shiga, and H. Kawai, “An investigation of subband wavenet vocoder covering entire audible frequency range with limited acoustic features,” in IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , pp. 5654–565...

    2018

  23. [31]

    Multirate digital signal processing,

    R. E. Crociere and L. R. Rabiner, “Multirate digital signal processing,” Prentice Hall, Englewood Cliffs, 1983. 10 DurIAN: Duration Informed Attention Network For Multimodal Synthesis Appendix A Filter Design For Multiband WaveRNN A stable yet more efficient low cost filter bank...

    1983

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.