티스토리 뷰

반응형

STYLER: Style Factor Modeling with Rapidity and Robustness via Speech Decomposition for Expressive and Controllable Neural Text to Speech


  • Text-to-Speech는 어려운 합성 condition에 대한 robustness와 expressiveness, controllability를 요구함
  • STYLER
    • Mel-Calibrator를 통한 audio-text aligning을 도입하여 unseen data에 대한 robust 한 추론을 가능하게 함
    • Supervision 하에서 disentangled style factor modeling을 통해 controllability를 향상
    • Domain adversarial training, Residual decoding을 활용하여 noise-robust style transfer를 지원하고 additional label 없이 noise를 decompose
  • 논문 (INTERSPEECH 2021) : Paper Link

1. Introduction

  • Text-to-Speech (TTS)는 많은 발전이 있었지만 여전히 expressiveness와 controllability가 부족함
    • 합성된 음성의 style은 training data에 대한 평균적인 style로 결정되므로 TTS의 expressiveness가 제한됨
    • BUT, 이를 해결하기 위한 expressive TTS 모델은 autoregressive architecture로 인해 속도와 robustness의 한계가 존재
      - 각 frame은 예측을 위해 이전의 모든 time step을 반복해야 하기 때문에 상당한 overhead가 발생
      - Unsupervised style modeling은 학습이 어렵고 feature를 disentangle 하기 어려움
    • Non-autoregressive 방식은 이러한 문제를 해결할 수 있음
      - Autoregressive decoder를 supervision이 있는 duration predictor로 대체하면 stability가 향상되고 속도도 빨라짐
      - BUT, 이러한 방식도 single text가 아닌 input에 대한 expressiveness와 contorllability는 부족
    • 결과적으로 TTS 모델에서 속도와 robustness, expressiveness, controllability를 모두 만족하기 위해서는,
      1. 속도와 robustness에 약점이 있는 autoregressive architecture를 대체해야 함
      2. Expressiveness, controllability를 위해 style factor modeling이 필요
        - Single text 외의 input variation을 제공

-> 그래서 빠르고 robust 하면서 expressive, controllable 한 TTS를 지원하는 STYLER를 제안 

 

  • STYLER
    • Non-autoregressive decoding시 각 style factor를 control 하는 style factor modeling을 도입
    • Source audio를 text input 외에도 duration, pitch, energy, content, noise의 5개 component로 split
      - 각 style factor를 개별적으로 encoding 한 다음, 음성으로 decoding
    • Noise modeling은 label 없이 reference의 noise를 decompose

< Overall of STYLER >

  • Audio-text aligning을 통해 unseen data에서 더 빠르고 robust 한 합성이 가능
  • Disentangled style factor를 통해 높은 expressivity와 controllability를 달성
  • Noise modeling pipeline을 통해 additional label 없이 noise를 style factor로 decompose 가능

2. Method

- Supervised Speech Decomposition

  • STYLER는 Style factor modeling과 Non-autoregressive TTS, 2가지에 대해 고려
    • Voice conversion과 달리 output speech의 content는 text이므로 audio content와 match 되지 않을 수 있음
    • 최신 non-autoregressive 모델은 supervision을 활용할 수 있지만, speech decomposition은 unsupervised condition에서 수행됨
    • 결과적으로 이를 위해 STYLER는 Mel Calibrator를 도입하고 supervision 하에서의 speech decomposition을 수행
  • Mel Calibrator
    • Text와 audio의 length mismatch는 attention mechanism을 통해 해결될 수 있음
      - BUT, unseen data에 대해서는 unstable 하다는 단점이 있음
    • 따라서 unseen data에 대한 robustness를 향상하기 위해 Mel Calibrator를 도입
      - Mel Calibrator는 input text와 audio의 total length를 사용하여 frame의 평균을 구하거나 frame을 반복하여,
      - Audio frame length에 대한 phoneme sequence size의 raito를 기준으로 single phoneme에 할당
    • 이러한 scaling 과정에서 aligning process는 frame-wise bottleneck이 되므로 attention이나 forced alignment의 사용이 필요 없음
      - 결과적으로 Mel Calibrator를 사용하면 non-autoregressive에 대한 robustness 문제를 해결하고, audio-related style이 text에 expose 되지 않음
  • Information Bottleneck under Supervision
    • Unsupervised speech decomposition은 source audio에서 필요한 information을 전달하기 위해 channel-wise, frame-wise 모두에 대한 bottleneck을 필요로 함
      - 결과적으로 TTS 모델의 학습을 어렵고 time-consuming 하게 만듦
    • Supervision 하에서 decomposition을 적용하면 pitch contour나 energy 같은 pre-obtained feature를 활용할 수 있음
      - 이때 forced feature selection을 통한 과도한 regularization은 information shortage와 성능 저하를 발생시킴
      - 따라서, STYLER는 style factor modeling 중 supervision에 대한 information flow에 맞게 bottleneck을 적절하게 mitigate 함

- Model Architecture

  • STYLER는 text $Z_{t}$, duration $Z_{d}$, pitch $Z_{p}$, speaker $Z_{s}$, energy $Z_{e}$, noise $Z_{n}$ 6가지의 style factor를 활용함
    • Text는 다른 audio-related factor와 동일한 style factor로 취급하고,
    • Non-autoregressive TTS와 speech decomposition approach를 도입
  • Encoders
    • Text encoder는 2개의 Feed-Forward Transformer (FFT) block으로 구성
      - Phoneme sequence로 input text를 사용
      - 이때 text는 text encoder에 의해서만 encoding 되므로 audio content는 다른 encoder의 bottleneck을 통해 제거됨
    • Duration, pitch, energy, noise encoder는 group normalization이 포함된 3개의 $5 \times 1$ convolution layer, 2개의 bidirectional LSTM을 통한 bottleneck으로 구성
      - Mel Calibrator는 convolution stack 다음에 추가
      - Encoder output은 linear layer로 예측하기 전에 ReLU activation을 통하여 channel-wise upsample 되고 duration predictor의 경우, frame-wise upsampling이 적용됨
    • Duration, noise encoder는 mel-spectrogram을 input으로 사용하고, pitch encoder는 speaker normalized pitch contour (mean = 0.5, std = 0.25)를 사용
      - Energy encoder는 0에서 1로 scale 된 energy를 사용
    • Speaker encoding은 pre-trained speaker embedding으로부터 얻어짐
      - 이를 위해 DeepSpeaker 모델을 활용
  • Decoders
    • STYLER에는 duration, pitch, energy에 대한 3개의 predictor가 존재
      - Predictor는 latent code가 아닌 real value를 예측하므로 모든 information을 사용해야 함
      - 이를 위해 input에는 해당하는 encoding과 text encoding의 합이 포함되고, 이때 text encoding은 dependency를 맞추기 위해 4차원 공간으로 projection 됨
    • Pitch predictor는 최종적인 pitch contour를 예측하기 위해 speaker encoding을 추가적으로 사용
      - 이때 downsampled / upsampled pitch encoding 모두에 speaker encoding을 추가하면 speaker identity decomposition 성능이 향상됨
    • Disentangled style factor로부터 최종적인 mel-spectrogram을 예측하는 decoder는 4개의 FFT block으로 구성됨
      - 해당 decoder의 input은 text encoding, pitch predictor ouptut, energy predictor output, speaker encoding으로 구성

STYLER Architecture

- Noise Modeling

  • STYLER의 noise robustness를 향상하기 위해 additional label 없이 audio에서 noise를 decompose 할 수 있음
    • STYLER에서 noise는 모델 encoder에서 explicitly encoding 된 remaining factor 중 하나이고, 다른 style factor를 제외한 residual property를 기반으로 정의할 수 있음
    • 이에 따라 noise를 modeling 하기 위해서는, 다른 encoder가 noisy input에 대해서 noise information을 포함하지 않도록 해야 함
  • Domain Adversarial Training
    • Noise information을 포함하지 않는다는 것은 noise-independent feature를 추출한다는 것과 동일함
      - 이를 위해 Domain Adversarial Training (DAT)를 활용할 수 있음
    • 따라서 STYLER는 jointly trainable 한 augmentation label과 Gradient Reversal Layer (GRL)을 도입
      - 이때 각 predictor의 label은 class label의 역할을 수행
      - Noise encoder를 제외한 모든 audio-related encoding에는 DAT가 적용됨
      - GRL을 통과한 각 encoding은 augmentation classifier를 통해 augmentation posterior (original / augmented)를 예측하는 데 사용됨
    • Augmentation classifier는 256 hidden size의 2개의 fully-connected layer로 구성되고, layer noramlization, ReLU activation이 사용됨
      - Noise encoder를 제외한 각 encoder는 noise-independent 하므로 clean label과 비교됨
  • Residual Decoding
    • Residual decoding은 clean decoding, noisy decoding 두 phase로 구성
      1. Clean decoding은:
        모든 noise-independent encoding을 사용하여 clean mel-spectrogram을 예측
      2. Noisy decoding은:
        Noise encoder output을 noise-independent encoding에 추가하여 noise가 있는 mel-spectrogram을 예측
    • Noisy decoding에서는 noise encoder만 업데이트하면 되기 때문에, gradient가 다른 encoder를 통해 흐르지 않음
    • 결과적으로 residual decoding은 explicit label 없이 audio의 leftover part에 noise가 집중되도록 하는 implicit supervision으로 볼 수 있음

- Loss Calculation

  • Noise modeling이 없는 모델의 total loss는:
    $Loss_{clean} = l_{mel-clean} + l_{duration} + l_{pitch} + l_{energy}$
    - $l_{mel-clean}$ : 예측된 mel-spectrogram과 target 사이의 mean squared error
    - $l_{duration}, l_{pitch}, l_{energy}$ : FastSpeech2와 동일한 predictor loss
  • Noise modeling을 포함한 모델의 total loss는:
    $Loss_{total} = Loss_{clean} + l_{mel-noisy} + l_{aug}$
    - $l_{mel-noisy}$ : $l_{mel-clean}$과 동일하게 계산된 noisy decoding 값
    - $l_{aug}$ : augmentation classifier loss

3. Experiments

- Settings

  • Dataset : VCTK, WHAM!
  • Comparisons : FastSpeech2, Mellotron

- Results

  • Rapidity and Naturalness
    • STYLER는 style factor modeling으로 인해 FastSpeech2에 비해서는 다소 느린 속도를 보였지만, Mellotron에 비해 3.49배의 학습 속도와 8.96배의 추론 속도 향상을 보임
    • MOS 측면에서 STYLER는 Unseen (US), Non-parallel (NP) condition에서 비교 모델들보다 더 나은 naturalness를 보임

모델 성능 비교

  • Style Transfer
    • STYLER는 CMOS 측면에서도 FastSpeech2, Mellotron 보다 더 나은 expressiveness를 보임
    • 특히 STYLER는 Unseen (US), Non-Parallel (NP) 환경에서 성능이  크게 저하되지 않아 robust 함

Style Transfer 성능 비교

  • Style Factor Modeling
    • 합성된 audio의 mel-spectrogram을 비교해 보면, STYLER는 noisy reference audio에서도 clean audio를 생성할 수 있음
    • 특히 noise modeling의 경우 input audio에 대한 다양한 noise level을 예측하는 것으로 나타남
      - Noise-independent style이 무시되면 해당하는 encoding은 fixed style로 나타남
      - Text encoder가 비활성 되면, audio-related style factor는 올바르게 사용되어도 unintelligible한 음성을 합성함
    • 결과적으로 STYLER에 사용된 noise modeling은 reference audio의 noise를 decompose 함

Style Factor Decomposition 비교

 

반응형
댓글
최근에 올라온 글
최근에 달린 댓글
«   2024/09   »
1 2 3 4 5 6 7
8 9 10 11 12 13 14
15 16 17 18 19 20 21
22 23 24 25 26 27 28
29 30
Total
Today
Yesterday