티스토리 뷰

반응형

ALO-VC: Any-to-Any Low-Latency One-Shot Voice Conversion


  • Non-parallel low-latency one-shot phonetic posteriorgrams-based voice conversion을 통해 빠른 합성이 가능함
  • ALO-VC
    • Pre-trained speaker encoder, pitch predictor, positional encoding을 결합해 구성됨
    • ALO-VC-R은 pre-trained d-vector speaker encoder를 활용하고 ALO-VC-E는 ECAPA-TDNN을 활용해 성능을 개선
  • 논문 (INTERSPEECH 2023) : Paper Link

1. Introduction

  • Voice Conversion (VC)는 linguistic content를 유지하면서 source speaker의 utterance를 target speaker의 voice로 변환하는 것을 목표로 함
    • 특히 any-to-any one-shot VC system은 하나의 reference utterance 만을 사용하여 any target speaker에 대한 conversion을 수행할 수 있어야 함
      1. VQMIVC는 content representation에 대한 vector quantization (VQ)과 mutual information (MI)를 도입함
      2. DiffVC의 경우 diffusion-based decoder를 활용하여 unseen speaker에 대한 similarity를 향상함
    • BUT, 이러한 기존의 non-causal VC system은 많은 parameter를 요구하고 상당한 future frame look-ahead를 사용하므로 real-time application에서 latency 요구사항을 초과함
      - 이를 위해 FastVC, Streamable VQMIVC 등을 고려할 수 있지만, 합성 품질 측면에서 한계가 있음
    • 한편으로 Phonetic Posteriorgram (PPG) representation을 활용하면 고품질 causal system을 구축할 수 있음

-> 그래서 PPG representation을 기반으로 한 low-latency one-shot VC system인 ALO-VC를 제안

 

  • ALO-VC
    • Pitch predictor와 Causal positional encoding을 통합하여 PPG에 phoneme location information을 제공
    • 47.5ms의 look-ahead를 사용하여 low-latency one-shot VC를 수행하는 2가지 variant를 구축:
      1. ALO-VC-R은 pre-trained d-vector speaker encoder를 활용
      2. ALO-VC-E는 ECAPA-TDNN speaker encoder를 활용해 성능을 개선

< Overall of ALO-VC >

  • PPG를 causal content representation으로 채택하고 pitch predictor를 통해 pitch variance information을 반영
  • 추가적으로 pre-trained speaker encoder를 활용하여 similarity를 개선
  • 결과적으로 기존보다 뛰어난 합성 품질과 latency를 달성

2. Method

  • ALO-VC는 acoustic model, pitch extractor, speaker encoder, conversion model, LPCNet의 5가지 component로 구성됨

Overall of ALO-VC

- Acoustic Model

  • Acoustic model은 source utterance의 speaker independent linguistic content representation을 추출하는 것을 목표로 함
    • 이를 위해 논문에서는 Phonetic Posteriorgram (PPG)를 사용하고, 구조적으로는 Conformer architecture에 기반한 architecture를 활용
      - 이때 multi-head attention은 latency를 증가시키므로 사용되지 않음
    • 전체적으로 ALO-VC의 acoustic model은 convolution module, causal convolution module, causal conformer block, uni-directional LSTM layer로 구성됨 
      1. $3\times 3$ kernel을 가지는 single 1D convolution layer로 구성된 convolution module은 adjacent frame 간의 local correlation을 detect 하는 역할을 수행
      2. 해당 module 다음에는 ReLU activation이 포함된 causal convolution layer, 4개의 causal Conformer block, 2개의 uni-directional LSTM, softmax activation이 포함된 time-distributed fully-connected layer가 이어짐
      3. Causal conformer block에서는 causality를 보장하기 위해 각 time frame에 대해 independent하게 layer normalization을 적용함
      4. 최종적으로 512-dimensional PPG representation은 LSTM layer를 통해 추출됨
    • Acoustic model은 2.7M의 parameter를 가지고 10ms future look-ahead frame을 사용함
    • Training 시 ground-truth phoneme alignment와 acoustic model output 간의 cross-entropy로 최적화됨
      - 이때 PPG representation에서 source speaker identity가 expose되는 것을 방지하기 위해 conversion model과는 개별적으로 training 됨

- Speaker Encoder

  • Speaker embedding을 얻기 위해, Resemblyzer의 pre-trained d-vector와 SpeechBrain의 pre-trained ECAPA-TDNN을 활용 
    • d-vector speaker encoder는 VoxCeleb2로 train 되어 1.4M의 parameter를 가지고, ECAPA-TDNN은 VoxCeleb1, VoxCeleb2로 train되어 14.7M의 parameter를 가짐
    • 해당 pre-trained model을 통해 target speaker identity와 content representation을 separate 할 수 있음
      - 특히 speaker encoder는 acoustic/conversion model과 independent 하게 training 되므로 content information leakage를 방지 가능
    • 결과적으로 해당 speaker encoder를 기반으로 ALO-VC에 대한 2가지 variant를 구축할 수 있음:
      1. ALO-VC-R : d-vector speaker encoder를 사용
      2. ALO-VC-E : ECAPA-TDNN speaker encoder를 사용

- Conversion Model

  • Conversion model에서는 LPCNet을 통해 converted speech를 합성할 수 있도록 converted 18-dimensional Bark-scale cepstral coefficient (BFCC)와 converted pitch, voiced/unvoiced flag (VUF)를 생성함
    • 구조적으로는 pitch predictor, positional encoding module, time-distributed fully-connected layer, uni-directional LSTM layer로 구성됨 
      1. Pitch predictor는 4개의 1D causal convolution layer, 1개의 uni-directional LSTM으로 구성되고, converted pitch $F0_{trg}$와 VUF를 input으로 사용
        - 여기서 LPCNet의 bulit-in pitch detector를 source speech에서 $F0$와 VUF를 추출하는데 사용함
      2. $F0_{trg}$는 target speech의 log-scaled $F0$ statistics로 계산됨:
        (Eq. 1) $\log F0_{trg}=(\log F0_{src}-\mu_{src})*\left(\frac{\sigma_{trg}}{\sigma_{src}}\right)+ \mu_{trg}$
        - $\mu_{src}, \mu_{trg}$ : 각각 source speaker, target speaker의 $\log F0$ 평균
        - $\sigma_{src}, \sigma_{trg}$ : 각각 source speaker, target speaker의 $\log F0$ 분산
        - 해당 pitch conversion 이후, $\log F0_{trg}$는 $F0_{trg}$로 다시 변환됨
      3. 최종적으로 time-distributed fully-connected layer는 $F0_{trg}$와 VUF representation을 2D로 project 하고, pitch predictor는 VUF와 target speech prosody를 예측에 활용함
        - 이때 VUF와 pitch variance information은 PPG에 추가됨
    • Positional encoding은 PPG에서 phoneme location을 제공하기 위해 사용됨
      1. 이때 streaming inference를 위한 positional encoding은:
        (Eq. 2) $P(L+1,2i)=\sin\left(\frac{L+1}{10000^{2i/d}}\right), \,\,\, P(L+1,2i+1)=\cos\left(\frac{L+1}{10000^{2i/d}}\right)$
        - $L$ : previous input representation length,
        $d$ : output embedding space dimension
        - $i$ : output embedding space에서 dimension에 대한 position function을 mapping 하는 index ($0\leq i <d/2$)
      2. Streaming inference에서 $L$을 positional encoding으로 사용하면 future information이 사용되는 것을 방지 가능
    • 이후 output representation을 post-process 하기 위해 Tanh activation과 3개의 1D causal convolution layer로 구성된 PostNet을 도입함
      - 결과적으로 conversion model은 5.6M parameter를 가짐
    • Training 과정에서 conversion model은 ground-truth BSCC와 PostNet output representation 간의 Mean Absolute Error (MAE)를 통해 최적화됨
      - 이때 ground-truth pitch, predicted pitch 간의 MAE와 ground-truth VUF, predicted VUF 간의 MAE를 추가

- Fine-tuned LPCNet for each Gender

  • VC system의 성능을 더욱 향상하기 위해, 논문은 VCTK dataset의 male/female speaker에 대해 각각 fine-tuning 된 average LPCNet을 활용함
    - 이때 하나의 gender에 대해 0.0001의 learning rate로 base LPCNet을 training 함

3. Experiments

- Settings

  • Dataset : VCTK, LibriSpeech
  • Comparisons : VQMIVC, DiffVC

- Results

  • Subjective Test
    • Similarity 측면에서 ALO-VC는 가장 뛰어난 성능을 보임

Subjective Test

  • Objective Test
    • Objective test에서도 ALO-VC가 가장 우수한 naturalness와 similarity를 달성함

Objective Test

  • Real-Time Factors and Latency
    • RTF, Latency 측면에서도 ALO-VC가 가장 뛰어남

RTF, Latency 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