[Paper 리뷰] CLAPSpeech: Learning Prosody from Text Context with Contrastive Language-Audio Pre-training
CLAPSpeech: Learning Prosody form Text Context with Contrastive Language-Audio Pre-trainingExpressive text-to-speech를 위한 masked token reconstruction은 prosody를 효과적으로 모델링하는 것이 어려움CLAPSpeech서로 다른 context에서 동일한 text token의 prosody variance를 explicitly learning 하는 cross-modal contrastive pre-training framework를 활용Encoder input과 contrastive loss를 설계하여 joint multi-modal space에서 text context와 해당 prosody..
[Paper 리뷰] STEN-TTS: Improving Zero-Shot Cross-Lingual Transfer for Multi-Lingual TTS with Style-Enhanced Normalization Diffusion Framework
STEN-TTS: Improving Zero-Shot Cross-Lingual Transfer for Multi-Lingual TTS with Style-Enhanced Normalization Diffusion FrameworkMultilingual text-to-speech는 주로 fine-tuning을 활용하거나 personal style을 추출하는데 중점을 둠STEN-TTS3초의 reference 만으로 multilingual 합성을 수행하고 style을 유지하는 Style-Enhanced Normalization (STEN)을 도입추가적으로 diffusion model에 해당 STEN module을 결합하여 style을 simulate 함논문 (INTERSPEECH 2023) : Paper Li..
[Paper 리뷰] PVAE-TTS: Adaptive Text-to-Speech via Progressive Style Adaptation
PVAE-TTS: Adaptive Text-to-Speech via Progressive Style AdaptationAdaptive text-to-speech는 limited data에서 speaking style을 학습하기 어렵기 때문에 새로운 speaker에 대한 합성 품질이 떨어짐PVAE-TTSStyle에 점진적으로 adapting 하면서 data를 생성하는 Progressive Variational AutoEncoder를 채택추가적으로 adaptiation 성능을 향상하기 위해 Dynamic Style Layer Normalization을 도입논문 (ICASSP 2022) : Paper Link1. IntroductionText-to-Speech (TTS) system을 training 하기 위..
[Paper 리뷰] VISinger2: High-Fidelity End-to-End Singing Voice Synthesis Enhanced by Digital Signal Processing Synthesizer
VISinger2: High-Fidelity End-to-End Singing Voice Synthesis Enhanced by Digital Signal Processing SynthesizerSinging Voice Synthesis에서 VISinger는 우수한 합성 성능을 달성했지만 다음의 한계점이 존재함- Text-to-Phase problem, Glitches, Low sampling rateVISinger2Digital signal processing synthesizer를 통해 VISinger의 latent representation $z$로부터 periodic/aperiodic signal을 생성- Phase information 없이 latent representation을 추출하도록 p..
[Paper 리뷰] RefineGAN: Universally Generating Waveform Better than Ground Truth with Highly Accurate Pitch and Intensity Responses
RefineGAN: Universally Generating Waveform Better than Ground Truth with Highly Accurate Pitch and Intensity ResponsesGenerative Adversarial Network-based waveform generation은 discriminator에 크게 의존함- 따라서 generation process에 uncertainty가 존재하고 pitch/intensity mismatch가 발생함RefineGANRobustness, pitch/intensity accuracy를 유지하기 위해 pitch-guided refine architecture를 구성추가적으로 training을 stabilize 하기 위해 multi..
[Paper 리뷰] EmoSphere-TTS: Emotional Style and Intensity Modeling via Spherical Emotion Vector for Controllable Emotional Text-to-Speech
EmoSphere-TTS: Emotional Style and Intensity Modeling via Spherical Emotion Vector for Controllable Emotional Text-to-SpeechEmotional text-to-speech는 pre-defined label로 제한되므로 emotion의 변화를 효과적으로 반영하지 못함EmoSphere-TTSEmotional style, intensity를 control 하는 spherical emotion vector를 채택Human annotation 없이 arousal, valence, dominance pseudo-label을 사용하여 Cartesian-spherical transformation을 통해 emotion을 모델..