yongxuUSTC / sednn

deep learning based speech enhancement using keras or pytorch, make it easy to use
http://staff.ustc.edu.cn/~jundu/The%20team/yongxu/demo/SE_DNN_taslp.html
334 stars 124 forks source link

模型輸入特徵問題請教 #13

Open JohnChen1124 opened 6 years ago

JohnChen1124 commented 6 years ago

您好: 非常感謝你們所完成的程式,希望能抽空回答一個小小的問題。模型輸入特徵取完頻譜大小之後有做絕對值,但好像沒有取平方,但在A Regression Approach to Speech Enhancement Based on Deep Neural Networks這篇paper中特徵為log-power spectral,這是有什麼原因嗎?感謝你們~~~

qiuqiangkong commented 6 years ago

您好!您看到的代码是论文之后复现的,我们发现绝对值比平方PESQ有0.1左右的提升。所以在代码中使用了绝对值loss。

Best wishes,

Qiuqiang


From: JohnChen1124 notifications@github.com Sent: 23 June 2018 11:07:35 To: yongxuUSTC/sednn Cc: Subscribed Subject: [yongxuUSTC/sednn] 模型”入特徵†–}ˆ教 (#13)

您好: 非常感x你‚ƒ所完成的程式,希望能抽空回答一‚€小小的†–}。模型”入特徵取完lV大小之後有做^Œχ担孟›]有取平方,但在A Regression Approach to Speech Enhancement Based on Deep Neural Networks@篇paper中特徵žlog-power spectral,@是有什N原因†?感x你‚ƒ~~~

― You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHubhttps://github.com/yongxuUSTC/sednn/issues/13, or mute the threadhttps://github.com/notifications/unsubscribe-auth/AMt5yT4IvzbJm4ReqatBS2CSkLF1jUR1ks5t_hNngaJpZM4U0uwB.

JohnChen1124 commented 6 years ago

我了解了~所以代碼中的loss function是使用mean absolute error而非mean square error。

在擷取音檔特徵的過程中,將音檔轉換成spectrogram大小值之後有做絕對值,也有取log,但好像沒有取平方,是因為loss function改變而做的調整嗎?感謝您~

qiuqiangkong commented 6 years ago

Loss 是一个值得深入研究的方向。在代码中我们用了log Mel spectrogram,没有取平方。Loss可以选取Lp范数。当p=1时就是abs,当p=2时就是平方。可以尝试调。


From: JohnChen1124 notifications@github.com Sent: 23 June 2018 14:56:12 To: yongxuUSTC/sednn Cc: Kong Q Mr (PG/R - Elec Electronic Eng); Comment Subject: Re: [yongxuUSTC/sednn] 模型”入特徵†–}ˆ教 (#13)

我了解了~所以代a中的loss function是使用mean absolute error而非mean square error。

在”X取音™n特徵的^程中,Œ⒁™nD“Q成spectrogram大小值之後有做^Œχ担灿腥log,但好像›]有取平方,是因žloss function改ƒ而做的{整†?

― You are receiving this because you commented. Reply to this email directly, view it on GitHubhttps://github.com/yongxuUSTC/sednn/issues/13#issuecomment-399680167, or mute the threadhttps://github.com/notifications/unsubscribe-auth/AMt5yWlmzUM4kbl0KYdZLr8AArDhb7fNks5t_kj8gaJpZM4U0uwB.

Nickkk1124 commented 6 years ago

真的很感謝您的回覆,您說的log Mel spectrogram與Mel Filter Bank有關係嗎?log Mel spectrogram沒有取平方也是因為PESQ的測量結果嗎?

qiuqiangkong commented 6 years ago

不好意思说错了,我们用的是log spectrogram,不需要Mel。上一封邮件应为:

" Loss 是一个值得深入研究的方向。在代码中我们用了log spectrogram,没有取平方。Loss可以选取Lp范数。当p=1时就是abs,当p=2时就是平方。可以尝试调。 "


From: Nickkk1124 notifications@github.com Sent: 24 June 2018 07:18:16 To: yongxuUSTC/sednn Cc: Kong Q Mr (PG/R - Elec Electronic Eng); Comment Subject: Re: [yongxuUSTC/sednn] 模型”入特徵†–}ˆ教 (#13)

真的很感x您的回覆,您f的log Mel spectrogramcMel Filter Bank有P‚S†幔

― You are receiving this because you commented. Reply to this email directly, view it on GitHubhttps://github.com/yongxuUSTC/sednn/issues/13#issuecomment-399732638, or mute the threadhttps://github.com/notifications/unsubscribe-auth/AMt5ydGbRXaNL3LWueV4bCM2WVEhxpqxks5t_y8ogaJpZM4U0uwB.

Nickkk1124 commented 6 years ago

log Mel spectrogram沒有取平方也是因為PESQ的測量結果嗎?還是因為loss改變而跟著改變呢?感謝您~

qiuqiangkong commented 6 years ago

是的,是根据???果?用的。

Get Outlook for Androidhttps://aka.ms/ghei36


From: Nickkk1124 notifications@github.com Sent: Sunday, June 24, 2018 5:35:27 PM To: yongxuUSTC/sednn Cc: Kong Q Mr (PG/R - Elec Electronic Eng); Comment Subject: Re: [yongxuUSTC/sednn] 模型輸入特徵問題請教 (#13)

log Mel spectrogram沒有取平方也是因為PESQ的測量結果嗎?還是因為loss改變而跟著改變呢?感謝您~

— You are receiving this because you commented. Reply to this email directly, view it on GitHubhttps://github.com/yongxuUSTC/sednn/issues/13#issuecomment-399769272, or mute the threadhttps://github.com/notifications/unsubscribe-auth/AMt5yRHjKHn0MvXb0MzBJ9a_hrY6z-Izks5t_7_PgaJpZM4U0uwB.