エンドツーエンド音声合成に向けたNIIにおけるソフトウェア群 ～ TacotronとWaveNetのチュートリアル (Part 1)～
•
3 gostaram•1,940 visualizações
2019年1月27日(日)に金沢において開催された音声研究会(SP)で実施した［チュートリアル招待講演］エンドツーエンド音声合成に向けたNIIにおけるソフトウェア群 ～ TacotronとWaveNetのチュートリアル ～のスライドです。 発表者：シン ワン、安田裕介
Recomendados
Recomendados
Mais conteúdo relacionado
Mais procurados
Mais procurados (20)
Semelhante a エンドツーエンド音声合成に向けたNIIにおけるソフトウェア群 ～ TacotronとWaveNetのチュートリアル (Part 1)～
Semelhante a エンドツーエンド音声合成に向けたNIIにおけるソフトウェア群 ～ TacotronとWaveNetのチュートリアル (Part 1)～ (20)
Mais de Yamagishi Laboratory, National Institute of Informatics, Japan
Mais de Yamagishi Laboratory, National Institute of Informatics, Japan (11)
Último
Último (20)
エンドツーエンド音声合成に向けたNIIにおけるソフトウェア群 ～ TacotronとWaveNetのチュートリアル (Part 1)～
- 1. NII Tacotron WaveNet 1contact: wangxin@nii.ac.jp we welcome critical comments, suggestions, and discussion Xin WANG, Yusuke YASUDA National Institute of Informatics, Japan 2019-01-27
- 2. NII Part 1: WaveNet ssddddd 2contact: wangxin@nii.ac.jp we welcome cri4cal comments, sugges4ons, and discussion Xin WANG National Institute of Informatics, Japan 2019-01-27 Neural waveform models
- 3. l l l l PhD (2015-2018) l M.A (2012-2015) l h0p://tonywangx.github.io SELF-INTRODUCTION 3
- 4. l Introduction: text-to-speech synthesis l Neural waveform models l Summary & software CO N TEN TS 4
- 5. Text-to-speech synthesis (TTS) [1,2] NII’s TTS systems 5 INTRODUCTION [1] P. Taylor. Text-to-Speech Synthesis. Cambridge University Press, 2009. [2] T. Dutoit. An Introduction to Text-to-speech Synthesis. Kluwer Academic Publishers, Norwell, MA, USA, 1997. Text Text-to-speech (TTS) Speech waveform Text Waveform Text Waveform NII TTS system 2013 2016 2018 Year
- 6. 6 INTRODUCTION Text-to-speech synthesis (TTS) q Architectures text text-to-speech speechText Text analyzer Acous6c model(s) Linguistic features Waveform Waveform model Acoustic features … * | * || Linguistic features Acoustic features Spectral features,F0, Band-aperiodicity, etc.
- 7. 7 INTRODUCTION Text-to-speech synthesis (TTS) q Architectures TTS backend model text text-to-speech speechText Text analyzer Acoustic model(s) Linguistic features Waveform Waveform model Acoustic features Text Text analyzer Linguistic features Waveform End-to-end TTS modelText Waveform
- 8. 8 INTRODUCTION Text-to-speech synthesis (TTS) q Architectures TTS backend model sss text text-to-speech speechText Text analyzer Acoustic model(s) Linguistic features Waveform Waveform model Text Text analyzer Linguistic features Waveform End-to-end TTS modelText Waveform Waveform module Waveform module Tutorial part 1: neural waveform modeling Tutorial part 2: end-to-end TTS Acoustic features
- 9. l Introduction: text-to-speech synthesis l Neural waveform modeling • Overview • Autoregressive models • Normalizing flow • STFT-based training criterion l Summary CONTENTS 9
- 11. 11 OVERVIEW Task definition … c1 c2 cNc3 Waveform values Spectrogram 1 2 3 4 T… o1 o2 o3 o4 oT T waveform sampling points N frames
- 12. Neural waveform models 12 OVERVIEW Task definition 1 2 3 4 T… o1 o2 o3 o4 oT c1 … c2 cNc3 p(o1:T |c1:N ; ⇥) Waveform values
- 13. 13 OVERVIEW 1 2 3 4 T … c1 … c2 cNc3 Naïve model q Simple network + maximum-likelihood • Feedforward network Hidden layers Distribution p(o4|c1:N ; ⇥) …Up-sampling Waveform values
- 14. 14 OVERVIEW 1 2 3 4 T… c1 … c2 cNc3 Naïve model q Simple network + maximum-likelihood • RNN …Up-sampling Hidden layers Distribution Waveform values
- 15. 15 OVERVIEW 1 2 3 4 T… c1 … c2 cNc3 Naïve model q Simple network + maximum-likelihood • Dilated CNN [1,2] … Hidden layers Distribution Waveform values Up-sampling [1] F. Yu and V. Koltun. Multi-scale context aggregation by dilated convolutions. arXiv preprint arXiv:1511.07122, 2015. [2] A. Waibel, T. Hanazawa, G. Hinton, K. Shikano, and K. Lang. Phoneme recognition using time-delay neural networks. IEEE Transactions on Acoustics, Speech, and Signal Processing, 37(3):328–339, 1989.
- 16. 16 OVERVIEW 1 2 3 4 T… … Naïve model … c1 c2 cNc3 o1 o2 o3 o4 oT p(o1:T |c1:N ; ⇥) = p(o1|c1:N ; ⇥)· · ·p(oT |c1:N ; ⇥) = TY t=1 p(ot|c1:N ; ⇥) o1:TWeak temporal model <---------------> strongly correlated data mismatch Hidden layers Distribution Waveform values Up-sampling
- 17. 17 OVERVIEW Towards better models Naïve model o1:T c1:N Criterion Better model o1:T c1:N Criterion Better model Easier target Model o1:T c1:N Criterion Transform Model o1:T c1:N Criterion Better criterion Weak temporal model <---------------> strongly correlated data mismatch Combination Better Model o1:T c1:N Criterion Knowledge- based Transform
- 18. 18 OVERVIEW WaveNet WaveRNN SampleRNN FFTNet WaveGlow FloWaveNet Neural source- filter model Parallel WaveNet ClariNet RNN + STFT loss Model o1:T c1:N Criterion Model o1:T c1:N Criterion Transform Model o1:T c1:N Criterion Autoregressive models Normlizing flow / feature transformation Training criterion in frequency domain Universal neural vocoder Model o1:T c1:N Criterion Knowledge-based Transform Speech/signal processing ☛ Please check reference in the appendix LPCNet LP-WaveNet ExcitNet GlotNet Subband WaveNet
- 19. l Introduction: text-to-speech synthesis l Neural waveform models • Overview • Autoregressive models • Normalizing flow • STFT-based training criterion l Summary & software CONTENTS 19
- 20. 20 PART I: AUTOREGRESSIVE MODELS Core idea q Naïve model … … c1 c2 cNc3 p(o1:T |c1:N ; ⇥) = TY t=1 p(ot|c1:N ; ⇥) 1 2 3 4 T… o1 o2 o3 o4 oT
- 21. 21 PART I: AUTOREGRESSIVE MODELS Core idea q Autoregressive (AR) model 1 2 3 4 T… … … c1 c2 cNc3 o1 o2 o3 o4 oT 1 2 3 p(o1:T |c1:N ; ⇥) = TY t=1 p(ot|o1:t 1, c1:N ; ⇥)
- 22. 22 PART I: AUTOREGRESSIVE MODELS Core idea q Autoregressive (AR) model • Training: use natural waveform for feedback (teacher forcing [1]) 1 2 3 4 T… … … c1 c2 cNc3 o1 o2 o3 o4 oT 1 2 3 [1] R. J. Williams and D. Zipser. A learning algorithm for continually running fully recurrent neural networks. Neural computation, 1(2):270–280, 1989.
- 23. 23 PART I: AUTOREGRESSIVE MODELS Core idea q Autoregressive (AR) model • Training: use natural waveform for feedback (teacher forcing [1]) 1 2 3 4 T… … … c1 c2 cNc3 o1 o2 o3 o4 oT 1 2 3 [1] R. J. Williams and D. Zipser. A learning algorithm for continually running fully recurrent neural networks. Neural computation, 1(2):270–280, 1989. Natural waveform
- 24. 24 PART I: AUTOREGRESSIVE MODELS Core idea q Autoregressive (AR) model • Training: use natural waveform for feedback (teacher forcing [1]) • Generation: … … c1 c2 cNc3 1 2 1 bo1 2 bo2 3 bo3 4 3 bo4 T… boT p(bo1:T |c1:N ; ⇥) = QT t=1 p(bot|bot P :t 1, c1:N ; ⇥) [1] R. J. Williams and D. Zipser. A learning algorithm for continually running fully recurrent neural networks. Neural computation, 1(2):270–280, 1989. Generated waveform
- 25. LPCNet LP-WaveNet ExcitNet GlotNet Subband WaveNet SampleRNN 25 WaveRNN FFTNet WaveGlow FloWaveNet Neural source- filter model Parallel WaveNet ClariNet RNN + STFT loss Model o1:T c1:N Criterion Model o1:T c1:N Criterion Transform Model o1:T c1:N Criterion Normlizing flow / feature transformation Training criterion in frequency domain Universal neural vocoder Model o1:T c1:N Criterion Knowledge-based Transform Speech/signal processing PART I: AUTOREGRESSIVE MODELS WaveNet Autoregressive models
- 26. WaveNet q Network structure • Details: http://id.nii.ac.jp/1001/00185683/ http://tonywangx.github.io/pdfs/wavenet.pdf 26 PART I: AUTOREGRESSIVE MODELS 1 2 3 4 T 1 2 3 c1 c2 cNc3 A. van den 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.
- 27. SampleRNN 27 WaveRNN FFTNet LPCNet LP-WaveNet ExcitNet GlotNet WaveGlow FloWaveNet Neural source- filter model Parallel WaveNet ClariNet RNN + STFT loss Model o1:T c1:N Criterion Model o1:T c1:N Criterion Transform Model o1:T c1:N Criterion Normlizing flow / feature transformation Training criterion in frequency domain Subband WaveNet Universal neural vocoder Model o1:T c1:N Criterion Knowledge-based Transform PART I: AUTOREGRESSIVE MODELS WaveNet Autoregressive models ☛ Please check SampleRNN in the appendix q Issues with WaveNet & Sample RNN Generation is very very … very slow q Acceleration? • Engineering-based: parallelize/simplify computation • Knowledge-based: speech / signal processing theory Speech/signal processing
- 28. 28 PART I: AUTOREGRESSIVE MODELS WaveRNN q Linear time AR dependency in WaveNet q Subscale AR dependency + batch sampling 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 Waveform values 16 steps to generate
- 29. 29 PART I: AUTOREGRESSIVE MODELS WaveRNN q Linear time AR dependency in WaveNet q Subscale AR dependency + batch sampling • Generate multiple samples at the same time 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 16 steps to generate 10 steps to generate step1 step2 step3 …
- 30. 30 PART I: AUTOREGRESSIVE MODELS WaveRNN q Linear time AR dependency in WaveNet q Subscale AR dependency + batch sampling • Generate multiple samples at the same time 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 16 steps to generate 10 steps to generate step1 step2 step3 …
- 31. SampleRNN 31 WaveRNN FFTNet WaveGlow FloWaveNet Neural source- filter model Parallel WaveNet ClariNet RNN + STFT loss Model o1:T c1:N Criterion Model o1:T c1:N Criterion Transform Model o1:T c1:N Criterion Normlizing flow / feature transformation Training criterion in frequency domain Universal neural vocoder Model o1:T c1:N Criterion Knowledge-based Transform PART I: AUTOREGRESSIVE MODELS WaveNet Autoregressive models q AR models Generation is still slow p(o1:T |c1:N ; ⇥) = TY t=1 p(ot|o1:t 1, c1:N ; ⇥) Speech/signal processing LPCNet LP-WaveNet ExcitNet GlotNet Subband WaveNet
- 32. SampleRNN 32 WaveRNN FFTNet Model o1:T c1:N Criterion Model o1:T c1:N Criterion Transform Normlizing flow / feature transformation Universal neural vocoder Model o1:T c1:N Criterion Knowledge-based Transform PART I: AUTOREGRESSIVE MODELS WaveNet Autoregressive models Non-autoregressive model? Speech/signal processing LPCNet LP-WaveNet ExcitNet GlotNet Subband WaveNet
- 33. l Introduction: text-to-speech synthesis l Neural waveform models • Overview • Autoregressive models • Normalizing flow • STFT-based training criterion l Summary & software CONTENTS 33
- 34. 34 PART II: NORMALIZING FLOW-BASED MODELS General idea • o with strong temporal correla3on ---> z with weak temporal correla3on • Principle of changing random variable Model s Criterion ⇥ o f 1 (o) c Model s⇥ c bo po(o|c; ⇥, ) = pz(z = f 1 (o)|c; ⇥) det @f 1 (o) @o bz f (bz) z training generation ☛ https://en.wikipedia.org/wiki/Probability_density_function#Dependent_variables_and_change_of_variables
- 35. 35 PART II: NO RM ALIZIN G FLOW -BASED MODELSGeneral idea • o ---> Gaussian noise sequence z • Multiple transformations: normalizing flow o c bo … f 1 M (·) N(o; I) Criterion c … N(o; I) bz z po(o|c; 1, · · · , M ) = pz(z = f 1 1 · · · f 1 M 1 f 1 M (o)|c) MY m=1 det J(f 1 m ) f 1 1 (·) f 1 M 1 (·) f M (·) f M 1 (·) f 1 (·) training generation D. Rezende and S. Mohamed. Variational inference with normalizing flows. In Proc. ICML, pages 1530–1538, 2015.
- 36. SampleRNN 36 WaveRNN FFTNetLPCNet LP-WaveNet ExcitNet GlotNet Model o1:T c1:N Criterion Subband WaveNet Universal neural vocoder Model o1:T c1:N Criterion Knowledge-based Transform Speech production model WaveNet Autoregressive models Model o1:T c1:N Criterion Transform PART II: NORMALIZING FLOW-BASED MODELS po(o|c; 1, · · · , M ) = pz(z = f 1 1 · · · f 1 M 1 f 1 M (o)|c) MY m=1 det J(f 1 m ) WaveGlow FloWaveNet Parallel WaveNet ClariNet f m (·) in linear time domain f m (·) in subscale time domain f m (·)• Different time complexity , different training strategies 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. W. Ping, K. Peng, and J. Chen. Clarinet: Parallel wave generation in end-to-end text-to-speech. arXiv preprint arXiv:1807.07281, 2018. S. Kim, S.-g. Lee, J. Song, and S. Yoon. Flowavenet: A generative flow for raw audio. arXiv preprint arXiv:1811.02155, 2018. R. Prenger, R. Valle, and B. Catanzaro. Waveglow: A flow-based generative network for speech synthesis. arXiv preprint arXiv:1811.00002, 2018.
- 37. 37 PART II: NORMALIZING FLOW-BASED MODELS ClariNet & parallel WaveNet q Genera(on process … N(o; I) c1:N bo1:T f M (·) f M 1 (·) f 1 (·) CNN 1 (·) 1 2 3 4 T 1 2 3 4 T 1:T µ1:T bz (0) 1:T 1 2 3 4 T ü Parallel computation ü Fast generation * Initial condition may be implemented differently 1 2 3 4 T bz (0) 1:T bz (1) 1:T bz (1) t = bz (0) t · t + µt bz (1) 1:T
- 38. 38 PART II: NORMALIZING FLOW-BASED MODELS ClariNet & parallel WaveNet q Naïve training process … N(o; I) c1:N 1 2 3 4 T 1 1 2 2 3 3 4 4 T T CNN 1 (·) 1 2 3 4 1:T µ1:T T f 1 M (·) f 1 1 (·) f 1 M 1 (·) o1:T z (1) 1:T z (0) 1:T Training 7me ~ O(T) z (0) t = (z (1) t µt)/ t z (1) 1:T z (0) 1:T
- 39. 39 PART II: NORMALIZING FLOW-BASED MODELS o c … f 1 M (·) N(o; I) Criterion z f 1 1 (·) f 1 M 1 (·) Training Generation (Inverse-AR [1]) Normalizing flow bo c … N(o; I) bz f M (·) f M 1 (·) f 1 (·) Original WaveNet 1 2 3 4 1 2 3 1 2 3 4 1 2 3 ~ O(T) [1] D. P. Kingma, T. Salimans, R. Jozefowicz, X. Chen, I. Sutskever, and M. Welling. Improved variational inference with inverse autoregressive flow. In Proc. NIPS, pages 4743–4751, 2016. ✓ ~ O(1) ✓ ~ O(1) ~ O(T)
- 40. 40 PART II: NORMALIZING FLOW-BASED MODELS ClariNet & parallel WaveNet q Fast training : knowledge distilling • Teacher gives • Student gives … N(o; I) f M (·) f M 1 (·) f 1 (·) c1:N bo1:T 1 2 3 p(bot|z1:T , c1:T , student) p(bot|bo1:t 1, c1:T , teacher) Student learns from teacher Teacher Pre-trained WaveNet Student
- 41. SampleRNN 41 WaveRNN FFTNet Model o1:T c1:N Criterion Model o1:T c1:N Criterion Transform Normlizing flow / feature transformation Universal neural vocoder Model o1:T c1:N Criterion Knowledge-based Transform PART II: NORMALIZING FLOW-BASED MODELS WaveNet Autoregressive models Parallel WaveNet ClariNet q Parallel WaveNet & ClariNet ü No AR dependency in generation Need pre-trained WaveNet Without teacher WaveNet ? Speech/signal processing LPCNet LP-WaveNet ExcitNet GlotNet Subband WaveNet
- 42. WaveGlow q Why ~ O(T) ? Dependency in linear time domain q Reduce T? Dependency in subscale time domain 42 PART II: NORMALIZING FLOW-BASED MODELS f 1 1 (·) Generation Training 1 2 3 4 5 9 13 1 2 3 4 5 9 13 1 2 3 4 5 9 13 1 2 3 4 5 9 13 1 2 3 4 T 1 2 3 4 Generation 1 2 3 4 T 1 2 3 4 Training R. Prenger, R. Valle, and B. Catanzaro. Waveglow: A flow-based generative network for speech synthesis. arXiv preprint arXiv:1811.00002, 2018.
- 43. WaveGlow q Why ~ O(T) ? Dependency in linear time domain q Reduce T? Dependency in subscale time domain 43 PART II: NORMALIZING FLOW-BASED MODELS f 1 1 (·) Generation Training 1 2 3 4 5 9 13 1 2 3 4 5 9 13 1 2 3 4 5 9 13 1 2 3 4 5 9 13 1 2 3 4 T 1 2 3 4 Generation 1 2 3 4 T 1 2 3 4 Training R. Prenger, R. Valle, and B. Catanzaro. Waveglow: A flow-based generative network for speech synthesis. arXiv preprint arXiv:1811.00002, 2018. 1 2 3 4 5 9 13 1 2 3 4
- 44. SampleRNN 44 WaveRNN FFTNet Model o1:T c1:N Criterion Model o1:T c1:N Criterion Transform Normlizing flow / feature transformation Universal neural vocoder Model o1:T c1:N Criterion Knowledge-based Transform PART II: NORMALIZING FLOW-BASED MODELS WaveNet Autoregressive models Parallel WaveNet ClariNet WaveGlow FloWaveNet ?Simple model without normalizing flow Neural source- filter model Speech/signal processing LPCNet LP-WaveNet ExcitNet GlotNet Subband WaveNet
- 45. l Introduction: text-to-speech synthesis l Neural waveform models • Overview • Autoregressive models • Normalizing flow • STFT-based training criterion l Summary & software CONTENTS 45
- 46. 46 PART III: STFT-BASED TRAINING CRITERION 1 2 3 4 T… c1 … c2 cNc3 Neural source-filter model q Naïve model and STFT-based criterion … Generated waveforms X. Wang, S. Takaki, and J. Yamagishi. Neural source-filter-based waveform model for statistical parametric speech synthesis. arXiv preprint arXiv:1810.11946, 2018.
- 47. 47 PART III: STFT-BASED TRAINING CRITERION 1 2 3 4 T… c1 … c2 cNc3 Neural source-filter model q Naïve model and STFT-based criterion … Generated waveforms Natural waveforms STFT STFT Spectral amplitude error 1 2 3 4 T X. Wang, S. Takaki, and J. Yamagishi. Neural source-filter-based waveform model for statistical parametric speech synthesis. arXiv preprint arXiv:1810.11946, 2018.
- 48. 48 PART III: STFT-BASED TRAIN IN G CRITERIO N 1 2 3 4 T… c1 … c2 cNc3 Neural source-filter model q Naïve model and STFT-based criterion … Generated waveforms 1 2 3 4 T Natural waveforms STFT STFT Spectral amplitude error 1 2 3 4 T Sine & harmonics F0
- 49. 49 PART III: STFT-BASED TRAINING CRITERION Neural source-filter model q Examples Only harmonics No formants
- 50. 50 PART III: STFT-BASED TRAIN IN G CRITERIO N Neural source-filter model q Examples
- 51. 51 PART III: STFT-BASED TRAINING CRITERION Neural source-filter model q Examples
- 52. 52 PART III: STFT-BASED TRAINING CRITERION Neural source-filter model q Examples
- 53. 53 PART III: STFT-BASED TRAINING CRITERION Neural source-filter model q Examples
- 54. 54 PART III: STFT-BASED TRAINING CRITERION Neural source-filter model q Examples
- 55. 55 PART III: STFT-BASED TRAINING CRITERION Neural source-filter model q Results • Faster than WaveNet (at least 100 2mes) • Smaller than WaveNet • Speech quality is similar to WaveNet Natural WOR WAD WAC NSF 1 2 3 4 5 Quality(MOS) WORLD vocoder WaveNet u-law WaveNet Gaussian Neural source-filter Copy-synthesis Text-to-speech https://nii-yamagishilab.github.io/samples-nsf/index.html
- 56. l Introduction: text-to-speech synthesis l Neural waveform models l Summary & software CO N TEN TS 56
- 57. 57 WaveNet WaveRNN SampleRNN FFTNet WaveGlow FloWaveNet Neural source- filter model Parallel WaveNet ClariNet RNN + STFT loss Model o1:T c1:N Criterion Model o1:T c1:N Criterion Transform Model o1:T c1:N Criterion Universal neural vocoder Model o1:T c1:N Criterion Knowledge-based Transform SU M M ARY LPCNet LP-WaveNet ExcitNet GlotNet Subband WaveNet
- 58. 58 WaveNet WaveRNN SampleRNN FFTNet WaveGlow FloWaveNet Neural source- filter model Parallel WaveNet ClariNet RNN + STFT loss Model o1:T c1:N Criterion Model o1:T c1:N Criterion Transform Model o1:T c1:N Criterion Universal neural vocoder Model o1:T c1:N Criterion Knowledge-based Transform SOFTWARE NII neural network toolkit LPCNet LP-WaveNet ExcitNet GlotNet Subband WaveNet
- 59. NII neural network toolkit q Neural waveform models 1. Toolkit cores: https://github.com/nii-yamagishilab/project-CURRENNT-public.git 2. Toolkit scripts: https://github.com/nii-yamagishilab/project-CURRENNT-scripts 59 SOFTWARE
- 60. NII neural network toolkit q Useful slides • http://tonywangx.github.io/pdfs/CURRENNT_TUTORIAL.pdf • Other related slides http://tonywangx.github.io/slides.html 60 SO FTW ARE
- 61. 61 REFERENCE WaveNet: A. van den 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. SampleRNN: S. Mehri, K. Kumar, I. Gulrajani, R. Kumar, S. Jain, J. Sotelo, A. Courville, and Y. Bengio. Samplernn: An unconditional end-to-end neural audio generation model. arXiv preprint arXiv:1612.07837, 2016. WaveRNN: N. Kalchbrenner, E. Elsen, K. Simonyan, et.al. Efficient neural audio synthesis. In J. Dy and A. Krause, editors, Proc. ICML, volume 80 of Proceedings of Machine Learning Research, pages 2410–2419, 10–15 Jul 2018. FFTNet: Z. Jin, A. Finkelstein, G. J. Mysore, and J. Lu. FFTNet: A real-time speaker-dependent neural vocoder. In Proc. ICASSP, pages 2251–2255. IEEE, 2018. Universal vocoder: J. Lorenzo-Trueba, T. Drugman, J. Latorre, T. Merritt, B. Putrycz, and R. Barra-Chicote. Robust universal neural vocoding. arXiv preprint arXiv:1811.06292, 2018. Subband WaveNet: 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 Proc. ICASSP, pages 5654–5658. 2018. Parallel WaveNet: A. van den Oord, Y. Li, I. Babuschkin, et. al.. Parallel WaveNet: Fast high-fidelity speech synthesis. In Proc. ICML, pages 3918–3926, 2018. ClariNet: W. Ping, K. Peng, and J. Chen. Clarinet: Parallel wave generation in end-to-end text-to-speech. arXiv preprint arXiv:1807.07281, 2018. FlowWaveNet: S. Kim, S.-g. Lee, J. Song, and S. Yoon. Flowavenet: A generative flow for raw audio. arXiv preprint arXiv:1811.02155, 2018. WaveGlow: R. Prenger, R. Valle, and B. Catanzaro. Waveglow: A flow-based generative network for speech synthesis. arXiv preprint arXiv:1811.00002, 2018. RNN+STFT: S. Takaki, T. Nakashika, X. Wang, and J. Yamagishi. STFT spectral loss for training a neural speech waveform model. In Proc. ICASSP (submitted), 2018. NSF: X. Wang, S. Takaki, and J. Yamagishi. Neural source-filter-based waveform model for statistical para- metric speech synthesis. arXiv preprint arXiv:1810.11946, 2018. LP-WavNet: M.-J. Hwang, F. Soong, F. Xie, X. Wang, and H.-G. Kang. Lp-wavenet: Linear prediction-based wavenet speech synthesis. arXiv preprint arXiv:1811.11913, 2018. GlotNet: L. Juvela, V. Tsiaras, B. Bollepalli, M. Airaksinen, J. Yamagishi, and P. Alku. Speaker-independent raw waveform model for glottal excitation. arXiv preprint arXiv:1804.09593, 2018. ExcitNet: E. Song, K. Byun, and H.-G. Kang. Excitnet vocoder: A neural excitation model for parametric speech synthesis systems. arXiv preprint arXiv:1811.04769, 2018. LPCNet: J.-M. Valin and J. Skoglund. Lpcnet: Improving neural speech synthesis through linear prediction. arXiv preprint arXiv:1810.11846, 2018.
- 62. End of Part 1 62 Codes, scripts, slides http://nii-yamagishilab.github.io This work was partially supported by JST CREST Grant Number JPMJCR18A6, Japan and by MEXT KAKENHI Grant Numbers (16H06302, 16K16096, 17H04687, 18H04120, 18H04112, 18KT0051), Japan.
- 63. Text-to-speech (TTS) q Speech samples from NII’s TTS system 63 INTRODUCTION Text Waveform Text Waveform NII TTS system 2013 2016 2018 Year
- 64. 64 PART I: AUTOREGRESSIVE MODELS SampleRNN q Idea • Hierarchical / multi-resolution dependency S. Mehri, K. Kumar, I. Gulrajani, R. Kumar, S. Jain, J. Sotelo, A. Courville, and Y. Bengio. Samplernn: An unconditional end-to-end neural audio generation model. arXiv preprint arXiv:1612.07837, 2016.
- 65. 65 PART I: AUTOREGRESSIVE MODELS SampleRNN q Example network structure • R: time resolution increased by * 2 Tie 1 Tie 2 Tie 3
- 66. 66 PART I: AUTOREGRESSIVE MODELS SampleRNN q Example network structure • Training • R: 8me resolu8on increased by * 2 1 2 3 4 5 6 7 8 9 T Tie 1: R = 1 Tie 2: R = 2 Tie 3: R = 4 1 2 3 4 5 6 7 8 9 1 2 3 4 5 6 7 8 1 2 3 4 5 6 7 8 9
- 67. 67 PART I: AUTOREGRESSIVE MODELS SampleRNN q Example network structure • Generation process 1 2 3 4 5 6 7 8 1 2 3 4 5 6 7 8 9 Tie 1: R = 1 Tie 2: R = 2 Tie 3: R = 4 1 2 1 3 2 4 3 5 4 6 5 7 6 8 7 9 8 9 T
- 68. 68 PART I: AUTOREGRESSIVE MODELS WaveNet q Variants • u-Law discrete waveform ---> continuous-valued waveform § Mixture of logistic distribution [1] § GMM / Single-Gaussian [2] • Quantization noise shaping [3], related noise shaping method [4] [1] T. Salimans, A. Karpathy, X. Chen, and D. P. Kingma. Pixelcnn++: Improving the pixelcnn with discretized logistic mixture likelihood and other modifications. arXiv preprint arXiv:1701.05517, 2017. [2] W. Ping, K. Peng, and J. Chen. Clarinet: Parallel wave generation in end-to-end text-to-speech. arXiv preprint arXiv:1807.07281, 2018. [3] T. Yoshimura, K. Hashimoto, K. Oura, Y. Nankaku, and K. Tokuda. Mel-cepstrum-based quantization noise shaping applied to neural-network-based speech waveform synthesis. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 26(7):1173–1180, 2018. [4] K. Tachibana, T. Toda, Y. Shiga, and H. Kawai. An investigation of noise shaping with perceptual weighting for WaveNet-based speech generation. In Proc. ICASSP, pages 5664–5668. IEEE, 2018. 1 2 1 2 3 4 3 softmax 1 2 1 2 3 4 3 GMM/Gaussian
- 69. 69 PART I: AUTOREGRESSIVE MODELS WaveRNN q WaveNet is inefficient • Computa4on cost 1. Imprac4cal for 16 bit PCM (so?max of size 65536) 2. Very deep network (50 dilated CNN …) 3. … • Time latency 1. Genera4on 4me ~ O(waveform_length) 1 2 3 4 T 1 2 3
- 70. 70 PART I: AUTOREGRESSIVE MODELS WaveRNN q WaveRNN strategies • Computation cost 1. Impractical for 16 bit PCM (softmax of size 65536) 2. Very deep network (50 dilated CNN …) 3. ... • Time latency 1. Generation time ~ O(waveform_length) 1 2 3 4 T 1 2 3 Recurrent + feedfoward layers N. Kalchbrenner, E. Elsen, K. Simonyan, S. Noury, N. Casagrande, E. Lockhart, F. Stimberg, A. van den Oord, S. Dieleman, and K. Kavukcuoglu. Efficient neural audio synthesis. In J. Dy and A. Krause, editors, Proc. ICML, volume 80 of Proceedings of Machine Learning Research, pages 2410–2419, Stock- holmsm ̈assan, Stockholm Sweden, 10–15 Jul 2018. PMLR. Two-level softmax RNN + Feedforward Subscale dependency + batch