TRANSFERRING NEURAL SPEECH WAVEFORM SYNTHESIZERS TO MUSICAL

[zbller]

リンク
https://arxiv.org/pdf/1910.12381.pdf
どんなもの？
Transfering 3 neural waveform synthesizers to musical instrument sound generation in 3 ways.
先行研究と比べてどこがすごい？
Evaluating the ability of the synthesizers to generate instrument sound.
技術と手法のキモはどこ？
They selected WaveNet, WaveGlow, neural-source-filter (NSF) and trained them in 3ways: Training from scratch, Zero-shot adaptation, and Fine-tuning.
どうやって有効だと検証した？
They did MOS test and NSF with Fine-tuning was the highest. NSF also get high score with Training from scrath. This indicates NSF is appropriate for generating instrument sounds that has a regular harmonic structrue.
Zero-shot adaptation of WaveNet and NSF created noisy sounds, but WaveGlow generated higher quality sound. This reason may stem from the property of inverse-AR flow, which deserves further analysis.
議論はある？
Written in above.
次に読むべき論文