1 of 137

Finding Sparse Subnetworksfor Self-Supervised ASR and Speech Synthesis

Cheng-I Jeff Lai 1.2

with Yang Zhang 2*, Alexander H. Liu 1*, Shiyu Chang 2*, Erica Cooper 3*

Yi-Lun Liao 1, Yung-Sung Chuang 1, Kaizhi Qian 2, Sameer Khurana 1,

Junichi Yamagishi 3, David Cox 2, James Glass 1

1MIT CSAIL 2MIT-IBM Watson AI Lab 3National Institute of Informatics

PARP

https://people.csail.mit.edu/clai24/parp

TTS-Pruning

https://people.csail.mit.edu/clai24/prune-tts

2 of 137

Self-Supervised Learning Framework in Speech

2

Wav2vec 2.0: A Framework for Self-Supervised Learning of Speech Representations (Baevski et al. NeurIPS 2020)

Unsupervised Cross-lingual Representation Learning for Speech Recognition (Conneau et al. Interspeech 2021)

HuBERT: How much can a bad teacher benefit ASR pre-training? (Hsu et al. ICASSP 2021)

3 of 137

Self-Supervised Learning Framework in Speech

3

Wav2vec 2.0: A Framework for Self-Supervised Learning of Speech Representations (Baevski et al. NeurIPS 2020)

Unsupervised Cross-lingual Representation Learning for Speech Recognition (Conneau et al. Interspeech 2021)

HuBERT: How much can a bad teacher benefit ASR pre-training? (Hsu et al. ICASSP 2021)

4 of 137

Let’s Wind the Clock back 2 years

4

5 of 137

Year 2019: Speech SSL only as a proof of concept

5

Audio Word2Vec: Unsupervised Learning of Audio Segment Representations using Sequence-to-Sequence Autoencoder (Chung et al. Interspeech 2016)

Completely Unsupervised Phoneme Recognition by Adversarially Learning Mapping Relationships from Audio Embeddings (Liu et al. Interspeech 2018)

Completely Unsupervised Phoneme Recognition by a Generative Adversarial Network Harmonized with Iteratively Refined Hidden Markov Models (Chen et al. Interspeech 2019)

Speech2Vec: A Sequence-to-Sequence Framework for Learning Word Embeddings from Speech (Chung et al. Interspeech 2018)

An Unsupervised Autoregressive Model for Speech Representation Learning (Chung et al. Interspeech 2019)

6 of 137

Year 2020: Break-Through

6

Wav2vec 2.0: A Framework for Self-Supervised Learning of Speech Representations (Baevski et al. NeurIPS 2020)

Unsupervised Cross-lingual Representation Learning for Speech Recognition (Conneau et al. Interspeech 2021)

HuBERT: How much can a bad teacher benefit ASR pre-training? (Hsu et al. ICASSP 2021)

7 of 137

Year 2021: the Takeaway from SUPERB

7

SUPERB: Speech processing Universal PERformance Benchmark (Yang et al. Interspeech 2021)

8 of 137

Year 2021: the Takeaway from SUPERB

8

SUPERB: Speech processing Universal PERformance Benchmark (Yang et al. Interspeech 2021)

Scaling up representation module is necessary

regardless of the SSL objectives

9 of 137

Self-Supervised Learning Framework in Speech

9

Wav2vec 2.0: A Framework for Self-Supervised Learning of Speech Representations (Baevski et al. NeurIPS 2020)

Unsupervised Cross-lingual Representation Learning for Speech Recognition (Conneau et al. Interspeech 2021)

HuBERT: How much can a bad teacher benefit ASR pre-training? (Hsu et al. ICASSP 2021)

10 of 137

Subnetwork Discovery for Speech SSL

10

Wav2vec 2.0: A Framework for Self-Supervised Learning of Speech Representations (Baevski et al. NeurIPS 2020)

Unsupervised Cross-lingual Representation Learning for Speech Recognition (Conneau et al. Interspeech 2021)

HuBERT: How much can a bad teacher benefit ASR pre-training? (Hsu et al. ICASSP 2021)

11 of 137

Subnetwork Discovery for Speech SSL

11

Wav2vec 2.0: A Framework for Self-Supervised Learning of Speech Representations (Baevski et al. NeurIPS 2020)

Unsupervised Cross-lingual Representation Learning for Speech Recognition (Conneau et al. Interspeech 2021)

HuBERT: How much can a bad teacher benefit ASR pre-training? (Hsu et al. ICASSP 2021)

Low-resource scenarios

e.g. 10min/1h/10h paired data

12 of 137

Overarching Goal & Broader Impact

Develop an orthogonal approach to existing speech SSL studies that

    • Reduces architectural complexity via efficient pruning
    • Attains lower WER under the same low-resource ASR setting

12

13 of 137

Overarching Goal & Broader Impact

Develop an orthogonal approach to existing speech SSL studies that

    • Reduces architectural complexity via efficient pruning
    • Attains lower WER under the same low-resource ASR setting

Broader impact in

    • Extend modern-day speech technology to more under-explored spoken languages
    • Introduce a flexible pruning scheme to current/future speech SSL frameworks with minimal additional computational costs

13

14 of 137

Two Types of Subnetwork Discovery

14

The Lottery Ticket Hypothesis: Finding Sparse, Trainable Neural Networks (Frankle et al. ICLR 2019)

The Lottery Ticket Hypothesis for Pre-trained BERT Network (Chen et al. NeurIPS 2021)

15 of 137

Two Types of Subnetwork Discovery

15

The Lottery Ticket Hypothesis: Finding Sparse, Trainable Neural Networks (Frankle et al. ICLR 2019)

The Lottery Ticket Hypothesis for Pre-trained BERT Network (Chen et al. NeurIPS 2021)

16 of 137

Two Types of Subnetwork Discovery

16

The Lottery Ticket Hypothesis: Finding Sparse, Trainable Neural Networks (Frankle et al. ICLR 2019)

The Lottery Ticket Hypothesis for Pre-trained BERT Network (Chen et al. NeurIPS 2021)

17 of 137

Two Types of Subnetwork Discovery

17

The Lottery Ticket Hypothesis: Finding Sparse, Trainable Neural Networks (Frankle et al. ICLR 2019)

The Lottery Ticket Hypothesis for Pre-trained BERT Network (Chen et al. NeurIPS 2021)

18 of 137

Two Types of Subnetwork Discovery

18

The Lottery Ticket Hypothesis: Finding Sparse, Trainable Neural Networks (Frankle et al. ICLR 2019)

The Lottery Ticket Hypothesis for Pre-trained BERT Network (Chen et al. NeurIPS 2021)

19 of 137

Two Types of Subnetwork Discovery

19

The Lottery Ticket Hypothesis: Finding Sparse, Trainable Neural Networks (Frankle et al. ICLR 2019)

The Lottery Ticket Hypothesis for Pre-trained BERT Network (Chen et al. NeurIPS 2021)

One-Shot Magnitude Pruning (OMP)

20 of 137

Two Types of Subnetwork Discovery

20

The Lottery Ticket Hypothesis: Finding Sparse, Trainable Neural Networks (Frankle et al. ICLR 2019)

The Lottery Ticket Hypothesis for Pre-trained BERT Network (Chen et al. NeurIPS 2021)

multiple iterations:

Iterative Magnitude Pruning (IMP)

21 of 137

Two Types of Subnetwork Discovery

21

The Lottery Ticket Hypothesis: Finding Sparse, Trainable Neural Networks (Frankle et al. ICLR 2019)

The Lottery Ticket Hypothesis for Pre-trained BERT Network (Chen et al. NeurIPS 2021)

multiple iterations:

Iterative Magnitude Pruning (IMP)

Issue!!

each downstream finetuning requires 8/24 GPUs 🡪 computationally infeasible for many

22 of 137

Two Types of Subnetwork Discovery

22

The Lottery Ticket Hypothesis: Finding Sparse, Trainable Neural Networks (Frankle et al. ICLR 2019)

The Lottery Ticket Hypothesis for Pre-trained BERT Network (Chen et al. NeurIPS 2021)

23 of 137

Two Types of Subnetwork Discovery

23

The Lottery Ticket Hypothesis: Finding Sparse, Trainable Neural Networks (Frankle et al. ICLR 2019)

The Lottery Ticket Hypothesis for Pre-trained BERT Network (Chen et al. NeurIPS 2021)

Pruning mask obtained without target task finetuning

24 of 137

Two Types of Subnetwork Discovery

24

The Lottery Ticket Hypothesis: Finding Sparse, Trainable Neural Networks (Frankle et al. ICLR 2019)

The Lottery Ticket Hypothesis for Pre-trained BERT Network (Chen et al. NeurIPS 2021)

25 of 137

Two Types of Subnetwork Discovery

25

The Lottery Ticket Hypothesis: Finding Sparse, Trainable Neural Networks (Frankle et al. ICLR 2019)

The Lottery Ticket Hypothesis for Pre-trained BERT Network (Chen et al. NeurIPS 2021)

26 of 137

Two Types of Subnetwork Discovery

26

The Lottery Ticket Hypothesis: Finding Sparse, Trainable Neural Networks (Frankle et al. ICLR 2019)

The Lottery Ticket Hypothesis for Pre-trained BERT Network (Chen et al. NeurIPS 2021)

Magnitude Pruning at Initialization (MPI)

or Cross-lingual Mask Transfer

27 of 137

Two Types of Subnetwork Discovery

27

The Lottery Ticket Hypothesis: Finding Sparse, Trainable Neural Networks (Frankle et al. ICLR 2019)

The Lottery Ticket Hypothesis for Pre-trained BERT Network (Chen et al. NeurIPS 2021)

Computationally cheap

28 of 137

Two Types of Subnetwork Discovery

28

The Lottery Ticket Hypothesis: Finding Sparse, Trainable Neural Networks (Frankle et al. ICLR 2019)

The Lottery Ticket Hypothesis for Pre-trained BERT Network (Chen et al. NeurIPS 2021)

Issue!!

Worse performance than (Type I)

task-aware subnetwork discovery

29 of 137

Can we discover subnetworks at pretrained initialization with similar sparsity as (Type I) task-aware subnetwork discovery,

30 of 137

Can we discover subnetworks at pretrained initialization with similar sparsity as (Type I) task-aware subnetwork discovery,

while inducing the same computational cost as (Type II) task-agnostic subnetwork discovery?

31 of 137

Prune-Adjust-Re-Prune (PARP)

31

same as Task-Agnostic Subnetwork Discovery:

Pruning mask obtained without target task finetuning

32 of 137

Prune-Adjust-Re-Prune (PARP)

32

same as Task-Agnostic Subnetwork Discovery

33 of 137

Prune-Adjust-Re-Prune (PARP)

33

Dynamic Network Surgery for Efficient DNNs (Guo et al. NIPS 2016)

34 of 137

Prune-Adjust-Re-Prune (PARP)

34

Dynamic Network Surgery for Efficient DNNs (Guo et al. NIPS 2016)

35 of 137

Prune-Adjust-Re-Prune (PARP)

35

Dynamic Network Surgery for Efficient DNNs (Guo et al. NIPS 2016)

Pruned-out weights made updatable was already proposed in the past

36 of 137

Prune-Adjust-Re-Prune (PARP)

36

37 of 137

Prune-Adjust-Re-Prune (PARP)

37

final subnetwork is different from

the initial subnetwork

38 of 137

Prune-Adjust-Re-Prune (PARP)

38

Computationally cheap

Same/better sparsity as

Task-Aware Subnetwork Discovery

39 of 137

Why though?

39

40 of 137

A Surprising Observation

40

41 of 137

A Surprising Observation

41

Unsupervised Pretraining Transfers Well Across Languages (Riviere et al. ICASSP 2020)

Unsupervised Cross-lingual Representation Learning for Speech Recognition (Conneau et al. Interspeech 2021)

42 of 137

A Surprising Observation

42

Unsupervised Pretraining Transfers Well Across Languages (Riviere et al. ICASSP 2020)

Unsupervised Cross-lingual Representation Learning for Speech Recognition (Conneau et al. Interspeech 2021)

43 of 137

A Surprising Observation

43

Non-trivial, high IOUs

across spoken languages at any model scale / amount of supervision (10min, 1h, 10h) / sparsity level (10-90%)

Unsupervised Pretraining Transfers Well Across Languages (Riviere et al. ICASSP 2020)

Unsupervised Cross-lingual Representation Learning for Speech Recognition (Conneau et al. Interspeech 2021)

44 of 137

A Surprising Observation

44

Non-trivial, high IOUs

w/ SSL pretrained initialization

Unsupervised Pretraining Transfers Well Across Languages (Riviere et al. ICASSP 2020)

Unsupervised Cross-lingual Representation Learning for Speech Recognition (Conneau et al. Interspeech 2021)

45 of 137

Any (Type II) task-agnostic subnetwork is a sufficiently good initial subnetwork for PARP Step 1 due to the high IOUs.

46 of 137

Any (Type II) task-agnostic subnetwork is a sufficiently good initial subnetwork for PARP Step 1 due to the high IOUs.

Furthermore, PARP Step 2 merely makes minimal adjustment

to the initial subnetwork.

47 of 137

Low-Resource English ASR

47

Wav2vec 2.0: A Framework for Self-Supervised Learning of Speech Representations (Baevski et al. NeurIPS 2020)

PARP (black line) attains lower

WER than IMP/OMP/MPI

48 of 137

Low-Resource English ASR

48

Wav2vec 2.0: A Framework for Self-Supervised Learning of Speech Representations (Baevski et al. NeurIPS 2020)

w/o LM decoding,

PARP offers 10% lower WER than full wav2vec 2.0 base/large, whereas IMP/OMP/MPI don’t

49 of 137

Low-Resource English ASR

49

Wav2vec 2.0: A Framework for Self-Supervised Learning of Speech Representations (Baevski et al. NeurIPS 2020)

50 of 137

Computational Cost in # of SSL Finetuning

50

51 of 137

Low-Resource Multi-Lingual ASR

51

Unsupervised Pretraining Transfers Well Across Languages (Riviere et al. ICASSP 2020)

Unsupervised Cross-lingual Representation Learning for Speech Recognition (Conneau et al. Interspeech 2021)

52 of 137

52

53 of 137

53

54 of 137

54

55 of 137

55

56 of 137

Cross-Lingual Mask Transferability

56

57 of 137

Cross-Lingual Mask Transferability

57

58 of 137

Mask Transfer with Regular Finetuning

58

59 of 137

Mask Transfer with Regular Finetuning

59

mismatches are PER increase over the

same pair transfer

60 of 137

Mask Transfer with Regular Finetuning

60

Are similar spoken language more transferrable?

Spanish, French, Italian, Kyrgyz, Dutch are less transferrable to Russian

Swedish, Turkish, Tatar, Mandarin are more transferrable to Russian

61 of 137

Mask Transfer with PARP Step 2

61

62 of 137

Mask Transfer with PARP Step 2

62

PER degradation is over the same pair transfer is gone!

63 of 137

Mask Transfer with PARP Step 2

63

Random Pruning does not improve – PARP Step 2 indeed makes minimal adjustments

(see also Section 4.2 for more verifications)

64 of 137

Observation on Pre-Trained BERT/XLNet

64

Non-trivial, high IOUs

across GLUE tasks

65 of 137

Observation on Pre-Trained BERT/XLNet

65

Non-trivial, high IOUs

across GLUE tasks

66 of 137

Mask Transfer with Regular Finetuning/PARP

66

67 of 137

Mask Transfer with Regular Finetuning/PARP

67

performance degradation over the same pair transfer is much smaller

68 of 137

Mask Transfer with Regular Finetuning/PARP

68

69 of 137

LM-Decoding Sparse wav2vec 2.0

69

https://github.com/facebook/Ax

70 of 137

“Middle Contextualized Representations are more Valuable for Downstream ASR”

70

HuBERT: How much can a bad teacher benefit ASR pre-training? (Hsu et al. ICASSP 2021)

SUPERB: Speech processing Universal PERformance Benchmark (Yang et al. Interspeech 2021)

Unsupervised Speech Recognition (Baevski et al. NeurIPS 2021)

71 of 137

“Middle Contextualized Representations are more Valuable for Downstream ASR”

71

HuBERT: How much can a bad teacher benefit ASR pre-training? (Hsu et al. ICASSP 2021)

SUPERB: Speech processing Universal PERformance Benchmark (Yang et al. Interspeech 2021)

Unsupervised Speech Recognition (Baevski et al. NeurIPS 2021)

72 of 137

Pruned Weights Localization Across Layers

72

Consistent across 10 spoken languages and sparsities, See Appendix 18.

73 of 137

Pruned Weights Localization Across Layers

73

Pruning as an alternative means for probing!

as less important weights are discarded.

Low-Complexity Probing via Finding Subnetworks (Cao et al. NAACL 2021)

74 of 137

Cross-Task Mask Transfer on SUPERB

74

SUPERB: Speech processing Universal PERformance Benchmark (Yang et al. Interspeech 2021)

75 of 137

Cross-Task Mask Transfer on SUPERB

75

SUPERB: Speech processing Universal PERformance Benchmark (Yang et al. Interspeech 2021)

76 of 137

Cross-Task Mask Transfer on SUPERB

76

SUPERB: Speech processing Universal PERformance Benchmark (Yang et al. Interspeech 2021)

77 of 137

Cross-Task Mask Transfer on SUPERB

77

SUPERB: Speech processing Universal PERformance Benchmark (Yang et al. Interspeech 2021)

78 of 137

A Good Pruning Algorithm avoids pruning out the Important Weights in Pre-Trained Initializations

78

Dynamic Network Surgery for Efficient DNNs (Guo et al. NIPS 2016)

Identifying and Controlling Important Neurons in Neural Machine Translation (Bau et al. ICLR 2018)

SNIP: Single-Shot Network Pruning based on Connection Sensitivity (Lee et al. ICLR 2019)

Importance Estimation for Neural Network Pruning (Molchanov et al. CVPR 2019)

Analyzing Redundancy in Pretrained Transformer Models (Dalvi et al. EMNLP 2020)

Low-Complexity Probing via Finding Subnetworks (Cao et al. NAACL 2021)

79 of 137

A Good Pruning Algorithm avoids pruning out the Important Weights in Pre-Trained Initializations

79

80 of 137

A Good Pruning Algorithm avoids pruning out the Important Weights in Pre-Trained Initializations

80

81 of 137

A Good Pruning Algorithm avoids pruning out the Important Weights in Pre-Trained Initializations

81

82 of 137

On the Interplay between Sparsity, Naturalness, Intelligibility, and Prosody

On the Interplay Between Sparsity, Naturalness, Intelligibility, and Prosody in Speech Synthesis (Lai et al. arxiv 2021)

83 of 137

Out of 500 Neural TTS papers,

a few papers on Efficient TTS

83

A Survey on Neural Speech Synthesis (Tan et al. arxiv 2021)

84 of 137

Out of 500 Neural TTS papers,

a few papers on Efficient TTS,

0 papers on TTS pruning*

84

*Efficient Neural Audio Synthesis (Kalchbrenner et al. ICML 2018)

A Survey on Neural Speech Synthesis (Tan et al. arxiv 2021)

85 of 137

“We point out future research direction on neural TTS, mainly in two categories…

Efficient speech synthesis: how to reduce the cost of speech synthesis including the cost of collecting and labeling data, training and serving TTS models, etc.”

85

A Survey on Neural Speech Synthesis (Tan et al. arxiv 2021)

86 of 137

“We point out future research direction on neural TTS, mainly in two categories…

Efficient speech synthesis: how to reduce the cost of speech synthesis including the cost of collecting and labeling data, training and serving TTS models, etc.”

86

A Survey on Neural Speech Synthesis (Tan et al. arxiv 2021)

87 of 137

87

A Survey on Neural Speech Synthesis (Tan et al. arxiv 2021)

  • Data-Efficient TTS
  • Parameter-Efficient TTS
  • Energy-Efficient TTS

88 of 137

88

Natural TTS Synthesis by Conditioning WaveNet on Mel Spectrogram Predictions (Shen et al. ICASSP 2018)

Neural Speech Synthesis with Transformer Network (Li et al. AAAI 2019)

Parallel WaveGAN: A fast waveform generation model based on generative adversarial networks with multi-resolution spectrogram (Yamamoto et al. ICASSP 2020)

89 of 137

89

Natural TTS Synthesis by Conditioning WaveNet on Mel Spectrogram Predictions (Shen et al. ICASSP 2018)

Neural Speech Synthesis with Transformer Network (Li et al. AAAI 2019)

Parallel WaveGAN: A fast waveform generation model based on generative adversarial networks with multi-resolution spectrogram (Yamamoto et al. ICASSP 2020)

Module 1: Encoder

90 of 137

90

Natural TTS Synthesis by Conditioning WaveNet on Mel Spectrogram Predictions (Shen et al. ICASSP 2018)

Neural Speech Synthesis with Transformer Network (Li et al. AAAI 2019)

Parallel WaveGAN: A fast waveform generation model based on generative adversarial networks with multi-resolution spectrogram (Yamamoto et al. ICASSP 2020)

Module 2: Decoder

91 of 137

91

Natural TTS Synthesis by Conditioning WaveNet on Mel Spectrogram Predictions (Shen et al. ICASSP 2018)

Neural Speech Synthesis with Transformer Network (Li et al. AAAI 2019)

Parallel WaveGAN: A fast waveform generation model based on generative adversarial networks with multi-resolution spectrogram (Yamamoto et al. ICASSP 2020)

Module 3: Pre-Net

92 of 137

92

Natural TTS Synthesis by Conditioning WaveNet on Mel Spectrogram Predictions (Shen et al. ICASSP 2018)

Neural Speech Synthesis with Transformer Network (Li et al. AAAI 2019)

Parallel WaveGAN: A fast waveform generation model based on generative adversarial networks with multi-resolution spectrogram (Yamamoto et al. ICASSP 2020)

Module 4: Post-Net

93 of 137

93

Natural TTS Synthesis by Conditioning WaveNet on Mel Spectrogram Predictions (Shen et al. ICASSP 2018)

Neural Speech Synthesis with Transformer Network (Li et al. AAAI 2019)

Parallel WaveGAN: A fast waveform generation model based on generative adversarial networks with multi-resolution spectrogram (Yamamoto et al. ICASSP 2020)

Module 4: Post-Net

Module: Generator

94 of 137

Model Setup

94

95 of 137

Pruning Algorithms & Training

95

96 of 137

Pruning Algorithms & Training

96

97 of 137

Pruning Algorithms & Training

97

98 of 137

98

99 of 137

Pruned Model Evaluation

99

Naturalness

5-point scale MOS tests, 100 HITs/test

Intelligibility

synthetic WER via Google ASR API

Prosody

mean/std F0 estimation (Hz) and averaged utterance duration (sec.)

100 of 137

Demo

100

have made it difficult for the Treasury to maintain close and continuing supervision.

101 of 137

Demo

101

have made it difficult for the Treasury to maintain close and continuing supervision.

Unpruned

30% sparse

70% sparse

90% sparse

102 of 137

Demo

102

no sizable organization can achieve efficiency without the careful analysis and demarcation of responsibility

103 of 137

Demo

103

no sizable organization can achieve efficiency without the careful analysis and demarcation of responsibility

Unpruned

30% sparse

90% sparse

90% + 88% sparse

104 of 137

Does Sparsity Improve Naturalness?

104

105 of 137

Does Sparsity Improve Naturalness?

105

106 of 137

Does Sparsity Improve Naturalness?

106

No statistical difference up to 90% sparsity

107 of 137

Does Sparsity Improve Naturalness?

107

Hypothesis I: pruned model could train better

108 of 137

Does Sparsity Improve Naturalness?

108

No statistical difference up to 88% sparsity

109 of 137

Does Sparsity Improve Naturalness?

109

Hypothesis II: Neural TTS are over-parameterized in terms of Naturalness

110 of 137

Does Sparsity Improve Intelligibility?

110

111 of 137

Does Sparsity Improve Intelligibility?

111

112 of 137

Does Sparsity Improve Intelligibility?

112

1) Transformer-TTS: WER reduces prior to 90% 2) Tacotron2: WER maintains up to 40%

113 of 137

Does Sparsity Improve Intelligibility?

113

1) Transformer-TTS: WER reduces prior to 90% 2) Tacotron2: WER maintains up to 40%

3) P-WaveGAN: no WER change at all

114 of 137

Does Sparsity Improve Intelligibility?

114

Hypothesis III: Neural TTS are over-parameterized in terms of Intelligibility

115 of 137

Does Sparsity Change Prosody?

115

116 of 137

Does Sparsity Change Prosody?

116

117 of 137

117

1) Transformer-TTS: minimal duration fluctuation

2) Tacotron2: linear decrease (speedup)

118 of 137

118

1) Transformer-TTS: minimal duration fluctuation

2) Tacotron2: linear decrease (speedup)

3) P-WaveGAN: no duration change at all

119 of 137

119

Instability of Tacotron2 (RNN) pruning

c.f. Transformer-TTS F0 variation

120 of 137

120

Hypothesis IV: vocoder is not

responsible for prosody generation

121 of 137

121

Minimal F0 std variation c.f. ground truth 53 Hz F0 std

122 of 137

122

Minimal F0 over-smoothing

A Survey on Neural Speech Synthesis (Tan et al. arxiv 2021)

123 of 137

123

Hypothesis V: pruning does not hurt prosodic expressivity

124 of 137

Does more finetuning data improve Sparsity?

124

125 of 137

Does more finetuning data improve Sparsity?

125

30% of finetuning data (~7h) is enough

126 of 137

TTS-Augmentation for Unspoken Text

126

127 of 137

TTS-Augmentation for Unspoken Text

127

100h more data,

without domain mismatch

128 of 137

Does more finetuning data improve Sparsity?

128

TTS-Augmentation has minimal effects

129 of 137

129

130 of 137

130

131 of 137

Neural TTS models are

over-parameterized and highly prunable

132 of 137

Pruned models produce speech with

equal naturalness & intelligibility, with

similar prosody.

133 of 137

Pruning as an alternative tool for understanding deeper how

modern speech models are trained

134 of 137

Cheng-I Jeff Lai

Yang Zhang *

Alexander H. Liu *

Shiyu Chang *

Yi-Lun Liao

Yung-Sung Chuang

Kaizhi Qian

Sameer Khurana

David D. Cox

James R. Glass

Erica Cooper *

Junichi Yamagishi

135 of 137

Thank you!

PARP

https://people.csail.mit.edu/clai24/parp

TTS-Pruning

https://people.csail.mit.edu/clai24/prune-tts

136 of 137

Thank you!

PARP

https://people.csail.mit.edu/clai24/parp

TTS-Pruning

https://people.csail.mit.edu/clai24/prune-tts

137 of 137

Suggestions:

  • 1. stress it’s “weight” pruning.
  • 2. slide 31, re-prune 🡪 Unstructured Magnitude Pruning
  • 3. typo in figures: Unstructured “Magnitued” Pruning 🡪 Unstructured Magnitude Pruning
  • Jim & Andrew:
    • Break down computation i.e. which stage requires more compute
    • How transferrable results are between generalized?
      • Be more specific on the languages I did in the dataset
        • Go over related work
        • Longer talk to SLS?

137