Interspeech 2019 Voice Conversion Paper Review

Oct 10, 2019 • softrime

For Interspeech 2019, this year there are two sessions about Voice Conversion(VC). In this post, I would mostly review VC-related papers in session ’ Neural Techniques for Voice Conversion and Waveform Generation’, which is mainly about speaker information transformation. It is interesting that StarGAN becomes very popular this year. All the three papers about StarGAN tries to improve performace by modifying its architecture or training strategy. Also is One-shot Learning VC (three papers) which convert source speech to arbitrary target speaker with very limited target speaker corpus. One of them uses VAE while other two methods use PPG. There are also three VC works named on Tomoki Toda which all based on VAE framework. (W.I.P)


Summary

  Title Task Framework Author Affiliation
1 Non-Parallel Voice Conversion Using Weighted Generative Adversarial Networks Non-parallel; many-to-many StarGAN; WORLD Dipjyoti Paul University of Crete, Greece
2 One-Shot Voice Conversion by Separating Speaker and Content Representations with Instance Normalization Non-parallel; One-shot   Ju-chieh Chou; (Hung-yi Lee) National Taiwan University
3 One-Shot Voice Conversion with Global Speaker Embeddings Non-parallel; One-shot   Hui Lu; (Helen Meng) Tsinghua-CUHK
4 Non-Parallel Voice Conversion with Cyclic Variational Autoencoder Non-parallel; one-to-one   Patrick Lumban Tobing; (Tomoki Toda) Nagoya University
5 StarGAN-VC2: Rethinking Conditional Methods for StarGAN-Based Voice Conversion Non-parallel; many-to-many   Takuhiro Kaneko NTT
6 Fast Learning for Non-Parallel Many-to-Many Voice Conversion with Residual Star Generative Adversarial Networks Non-parallel; many-to-many   Shengkui Zhao Alibaba Group(Damo)
7 One-Shot Voice Conversion with Disentangled Representations by Leveraging Phonetic Posteriorgrams Non-parallel; One-shot   Seyed Hamidreza Mohammadi ObEN
8 Investigation of F0 Conditioning and Fully Convolutional Networks in Variational Autoencoder Based Voice Conversion Non-parallel; one-to-one   Wen-Chin Huang; (Tomoki Toda) Nagoya University
9 Robustness of Statistical Voice Conversion Based on Direct Waveform Modification Against Background Sounds Environment Robustness   Yusuke Kurita; (Tomoki Toda) Nagoya University
10 Jointly Trained Conversion Model and WaveNet Vocoder for Non-Parallel Voice Conversion Using Mel-Spectrograms and Phonetic Posteriorgrams Non-parallel; inf-to-one   Songxiang Liu; (Lifa Sun); (Helen Meng) CUHK
11 Group Latent Embedding for Vector Quantized Variational Autoencoder in Non-Parallel Voice Conversion Non-parallel; one-to-one   Shaojin Ding Texas A&M University
12 Semi-Supervised Voice Conversion with Amortized Variational Inference Semi-optimized; one-to-one   Cory Stephenson Intel AI Lab

Review

Non-Parallel Voice Conversion Using Weighted Generative Adversarial Networks

This paper modifies loss function in StarGAN. In detail, authors add a weight factor on the adversarial loss when update Generators, which means . The weight can be considered as the fake confidence for a sample. It reduces the loss of samples which is considered as fake with a high confidence by Discriminator.