search

temy13 / papers

my papers memo on issues
0 stars 0 forks source link

Audio Denoising with Deep Network Priors #1

Open temy13 opened 4 years ago

temy13 commented 4 years ago

URL

https://paperswithcode.com/paper/audio-denoising-with-deep-network-priors

PDF

https://arxiv.org/pdf/1904.07612v2.pdf

Method

https://github.com/mosheman5/DNP

Data

https://github.com/mosheman5/DNP/tree/master/wav_files

Reference

https://paperswithcode.com/task/audio-denoising

info

peper: 5

Nov 2019

temy13 commented 4 years ago

unsupervisedらしい

The method is completely unsupervised and only trains on the specific audio clip that is being denoised.

事前にモデルを作っておく必要がない => denoiseに恐ろし時間がかかる？

temy13 commented 4 years ago

冒頭で画像のdenoisingの話をするが、音声と画像の違いについて適当に話している

 Instead, our
method tracks the sequence of network outputs during training and observes its behavior
temy13 commented 4 years ago

実験

1) CSIG: Mean opinion score (MOS) predictor of signal distortion [20], (2) CBAK: MOS predictor of background-noise intrusiveness [20], (3) COVL: MOS predictor of overall signal quality [20], (4) PESQ: Perceptual evaluation of speech quality [21], and (5) SSNR: Segmental SNR [22].

MMSE-LSAがわりかし強い。が、こいつは冒頭にnoiseがある前提なのでないと提案が最強

temy13 commented 4 years ago

Method

ぶっちゃけわからんがコード転がってるし使えばいいんじゃないかな

We create a ran- dom input signal z of the same dimension as the noisy signal y = x + n (we assume an additive noise model, and the clean signal x and the noise n are unknown)

WaveUnet

we employ the CNN architecture known as the WaveUnet

temy13 commented 4 years ago

音声分野でWAVEUNETはよく使われるらしい

https://arxiv.org/pdf/1811.11307.pdf

temy13 commented 4 years ago

Wave-U-Net: A Multi-Scale Neural Network for End-to-End Audio Source Separation

https://arxiv.org/abs/1806.03185