論文解説: Tian (2020) FeatherWave: An efficient high-fidelity neural vocoder with multi-band linear prediction

提案モデル: mel-spec input (pitch-less), multiband LPCNet¹

デモ

中国語デモ
wavecoder.github.io

ConditioningNetwork

入力にmel-specをそのまま利用^{2, 3, 4}（pitch無し⁵, 80 dim⁶）。

Mel2LPcoeff

LP係数はmel-specから計算⁷。各バンドではmel-specのうち特定バンドのみからLP係数を算出⁸。

Original Paper

@misc{2005.05551,
Author = {Qiao Tian and Zewang Zhang and Heng Lu and Ling-Hui Chen and Shan Liu},
Title = {FeatherWave: An efficient high-fidelity neural vocoder with multi-band linear prediction},
Year = {2020},
Eprint = {arXiv:2005.05551},
}

---

1. “we merged multi-band into LPCNet framework which only conditioned on mel spectrograms.”↩

2. “It consists of a condition network that operates on input frames of mel spectrograms” from the paper↩

3. “only mel spectrograms, which are widely used in neural TTS systems, are adopted as input conditional features.” from the paper↩

4. “Since only mel-spectrograms are used in condition network” from the paper↩

5. “Since we use mel-spectrograms to extract the LP filters, the proposed model doesn’t depend on pitch extraction.”↩

6. “The 80 order melspectrograms were extracted as the conditions for all neural vocoders”↩

7. “the LP coefficients were estimated from the melspectrograms”↩

8. “M order linear prediction coefficients of each sub frequency band, α^b_k, can be extracted from the corresponding frequency bins of mel-spectrogram frame.”↩