다수 화자 음성 변환을 위한 VAE의 잠재 공간 disentanglement와 변환 경로 학습

다국어 초록 (Multilingual Abstract)

One of the most critical obstacles in voice conversion is the requirement of parallel training data, which contain the same linguistic content utterances spoken by different speakers. Collecting such parallel data is highly expensive process, therefore many works attempted to use non-parallel training data for voice conversion. One of such successful approaches is using cycle-consistent adversarial networks (CycleGAN), which utilize the cycle consistency loss. The major drawback of CycleGAN based methods, however, is that they can handle only one-to-one voice conversion from a source speaker to a target speaker, which makes it difficult to use for general-purpose cases requiring many-to-many voice conversion among multiple speakers. Another group of approaches using variational autoencoder (VAE) can handle many-to-many voice conversion, but their sound qualities are much lower than that of CycleGAN based methods. In this paper, we propose new methods of the conversion path training and disentangling latent space vector for many-to-many voice conversion and improving the sound qualities.

목차 (Table of Contents)

• 제 1 장 서론 1
• 제 2 장 관련 연구 4
• 2.1. Conditional Variational Autoencoder 4
• 2.2. Disentangled Latent Space 6
• 2.2.1. Speaker Classifier 6
• 2.3. Training Methods of Conversion Path 8
• 2.3.1. Auxiliary Classifier 8
• 2.3.2. Cross Generative Adverarial Network 10
• 제 3 장 제안하는 모델 13
• 3.1. Entropy of latent space vector 13
• 3.2. Cycle Consistency Loss 15
• 3.3. Multiple Decoders 17
• 3.4. Merge Models 19
• 제 4 장 실험 22
• 4.1. 학습 데이터 22
• 4.2. VAE 모델 구조 22
• 4.3. 파라메타 조정 23
• 4.4. 객관적 평가 25
• 4.5. 주관적 평가 41
• 제 5 장 결론 및 향후 과제 44
• 참고 문헌 45