[Paper 리뷰] StarGANv2-VC: A Diverse, Unsupervised, Non-parallel Framework for Natural-Sounding Voice Conversion
StarGANv2-VC: A Diverse, Unsupervised, Non-parallel Framework for Natural-Sounding Voice ConversionUnsupervised non-parallel many-to-many voice conversion을 위해 generative adversarial network를 활용할 수 있음StarGANv2-VCAdversarial source classifier loss와 perceptual loss를 결합하여 사용Style encoder를 통해 plain reading speech를 stylistic speech로 변환논문 (INTERSPEECH 2021) : Paper Link1. IntroductionVoice Conversion (..
[Paper 리뷰] Blow: A Single-Scale Hyperconditioned Flow for Non-Parallel Raw-Audio Voice Conversion
Blow: A Single-Scale Hyperconditioned Flow for Non-Parallel Raw-Audio Voice ConversionMany-to-Many voice conversion을 위해서는 non-parallel data를 활용할 수 있어야 함BlowHypernetwork conditioning과 single-scale normalizing flow를 활용Single speaker identifier를 사용하여 frame-by-frame으로 end-to-end training 됨논문 (NeurIPS 2019) : Paper Link1. IntroductionRaw audio는 intermediate representation을 사용하는 것보다 더 많은 model capacit..
[Paper 리뷰] FragmentVC: Any-to-Any Voice Conversion by End-to-End Extracting and Fusing Fine-Grained Voice Fragments with Attention
FragmentVC: Any-to-Any Voice Conversion by End-to-End Extracting and Fusing Fine-Grained Voice Fragments with AttentionAny-to-Any voice conversion은 unseen any speaker에 대해 voice conversion을 수행하는 것을 목표로 함FragmentVCWav2Vec 2.0을 통해 source speaker의 latent phonetic structure를 얻고 target speaker의 spectral feature를 log mel-spectrogram을 통해 얻음두 가지의 서로 다른 feature space를 two-stage training process를 통해 align ..
[Paper 리뷰] SEF-VC: Speaker Embedding Free Zero-Shot Voice Conversion with Cross Attention
SEF-VC: Speaker Embedding Free Zero-Shot Voice Conversion with Cross AttentionZero-shot voice conversion은 unseen target speaker로의 변환을 수행할 수 있지만, speaker similarity 측면에서 한계가 있음SEF-VCSpeaker embedding을 사용하지 않고 Position-Agnostic Cross-Attention을 도입하여 reference speech에서 speaker timbre를 학습이후 HuBERT semantic token으로부터 non-autoregressive 방식으로 waveform을 reconstruct논문 (ICASSP 2024) : Paper Link1. Introdu..
[Paper 리뷰] DRVC: A Framework of Any-to-Any Voice Conversion with Self-Supervised Learning
DRVC: A Framework of Any-to-Any Voice Conversion with Self-Supervised LearningAny-to-Any voice conversion은 training data에서 벗어난 source/target speaker에 대해 voice conversion을 수행하는 것을 목표로 함- BUT, 기존의 disentangle-based model은 speaker/content style information를 얻는 과정에서 untangle overlapping 문제가 발생함DRVC (Disentangled Representation Voice Conversion)Content encoder, timbre encoder, generator로 구성된 end-to-e..
[Paper 리뷰] VQMIVC: Quantization and Mutual Information-based Unsupervised Speech Representation Disentanglement for One-Shot Voice Conversion
VQMIVC: Quantization and Mutual Information-based Unsupervised Speech Representation Disentanglement for One-Shot Voice ConversionOne-shot voice conversion은 speech representation disentanglement를 통해 효과적으로 수행될 수 있음- BUT, 기존 방식은 speech representation 간의 correlation을 무시하므로 content information이 leakage 될 수 있음VQMIVCContent encoding 과정에서 vector quantization을 사용하고 training 중에 correlation metric으로써 mutu..