SSNR is not as high as in the paper

Question

SSNR is not as high as in the paper

YinyanLiu opened this issue 6 years ago · 7 comments

Hello,

I have trained the SEGAN by myself and also use the pretrained model.
I got the same metrics value of noisy as in the paper with SSNR=1.68
But when I evaluate the pretrained model, I only got SSNR=0.78, rather than 7.73 in the paper.

I have try different methods to calculate the SSNR and all got the low SSNR.

Anyone have the same issue? or have reappear the metrics in the paper?

Answer 1 · 2019-09-29T03:02:20.000Z

Hello,

I have the same problem as you have. I only have about 0.5.
If you have already solved this problem, I would like you to tell me the solution.

Answer 2 · 2019-10-23T15:21:20.000Z

@YinyanLiu How can I use a pre-trained model?

Answer 3 · 2020-07-30T18:16:18.000Z

Hello,

I have trained the SEGAN by myself and also use the pretrained model.
I got the same metrics value of noisy as in the paper with SSNR=1.68
But when I evaluate the pretrained model, I only got SSNR=0.78, rather than 7.73 in the paper.

I have try different methods to calculate the SSNR and all got the low SSNR.

Anyone have the same issue? or have reappear the metrics in the paper?

你好，我想请教一下你的pesq，csig这些指标是怎么算的，可以指点一下嘛，谢谢

Answer 4 · 2021-01-25T04:06:31.000Z

Hello,
I have trained the SEGAN by myself and also use the pretrained model.
I got the same metrics value of noisy as in the paper with SSNR=1.68
But when I evaluate the pretrained model, I only got SSNR=0.78, rather than 7.73 in the paper.
I have try different methods to calculate the SSNR and all got the low SSNR.
Anyone have the same issue? or have reappear the metrics in the paper?

你好，我想请教一下你的pesq，csig这些指标是怎么算的，可以指点一下嘛，谢谢

你好，这些指标你计算了吗，我刚刚接触语音增强，指标还不太懂怎么计算

Answer 5 · 2021-01-25T04:14:54.000Z

Hi baijin, 你好。我记得这篇paper的源代码是可以复现的。我18年复现过，不知道你是怎么找到我联系方式的呢？至于指标，我刚找了一下代码，见附件。这个代码应该也是我几年前写的，希望能有帮助。祝好， Yan From: zhangbaijin <notifications@github.com> Reply-To: santi-pdp/segan <reply@reply.github.com> Date: Monday, January 25, 2021 at 3:06 PM To: santi-pdp/segan <segan@noreply.github.com> Cc: YinyanLiu <yinyan.yana.liu@gmail.com>, Mention <mention@noreply.github.com> Subject: Re: [santi-pdp/segan] SSNR is not as high as in the paper (#62) Hello, I have trained the SEGAN by myself and also use the pretrained model. I got the same metrics value of noisy as in the paper with SSNR=1.68 But when I evaluate the pretrained model, I only got SSNR=0.78, rather than 7.73 in the paper. I have try different methods to calculate the SSNR and all got the low SSNR. Anyone have the same issue? or have reappear the metrics in the paper? 你好，我想请教一下你的pesq，csig这些指标是怎么算的，可以指点一下嘛，谢谢你好，这些指标你计算了吗，我刚刚接触语音增强，指标还不太懂怎么计算 — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub, or unsubscribe.

Answer 6 · 2021-01-25T04:20:14.000Z

非常感谢你的时间哈，我也是在github上看到的，不好意思附件好像没有文件，如果您方便的话可以发到SemiZxf@163.com ,十分感谢^^

…

------------------ 原始邮件 ------------------ 发件人: "notifications"<notifications@github.com>; 发送时间: 2021年1月25日(星期一) 中午12:15 收件人: "santi-pdp/segan"<segan@noreply.github.com>; 抄送: "张晓峰南京邮电大学"<512631874@qq.com>; "Comment"<comment@noreply.github.com>; 主题: Re: [santi-pdp/segan] SSNR is not as high as in the paper (#62) Hi baijin, 你好。我记得这篇paper的源代码是可以复现的。我18年复现过，不知道你是怎么找到我联系方式的呢？至于指标，我刚找了一下代码，见附件。这个代码应该也是我几年前写的，希望能有帮助。祝好， Yan From: zhangbaijin <notifications@github.com> Reply-To: santi-pdp/segan <reply@reply.github.com> Date: Monday, January 25, 2021 at 3:06 PM To: santi-pdp/segan <segan@noreply.github.com> Cc: YinyanLiu <yinyan.yana.liu@gmail.com>, Mention <mention@noreply.github.com> Subject: Re: [santi-pdp/segan] SSNR is not as high as in the paper (#62) Hello, I have trained the SEGAN by myself and also use the pretrained model. I got the same metrics value of noisy as in the paper with SSNR=1.68 But when I evaluate the pretrained model, I only got SSNR=0.78, rather than 7.73 in the paper. I have try different methods to calculate the SSNR and all got the low SSNR. Anyone have the same issue? or have reappear the metrics in the paper? 你好，我想请教一下你的pesq，csig这些指标是怎么算的，可以指点一下嘛，谢谢你好，这些指标你计算了吗，我刚刚接触语音增强，指标还不太懂怎么计算 — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub, or unsubscribe. — You are receiving this because you commented. Reply to this email directly, view it on GitHub, or unsubscribe.

Answer 7 · 2021-02-01T15:44:45.000Z

您好，打扰了。segan指标的代码可以也发我一份吗？我因为之前回复过所以看到了你们的邮件记录。谢谢您。

…

---Original--- From: "YinyanLiu"<notifications@github.com> Date: Mon, Jan 25, 2021 12:15 PM To: "santi-pdp/segan"<segan@noreply.github.com>; Cc: "Subscribed"<subscribed@noreply.github.com>; Subject: Re: [santi-pdp/segan] SSNR is not as high as in the paper (#62) Hi baijin, 你好。我记得这篇paper的源代码是可以复现的。我18年复现过，不知道你是怎么找到我联系方式的呢？至于指标，我刚找了一下代码，见附件。这个代码应该也是我几年前写的，希望能有帮助。祝好， Yan From: zhangbaijin <notifications@github.com> Reply-To: santi-pdp/segan <reply@reply.github.com> Date: Monday, January 25, 2021 at 3:06 PM To: santi-pdp/segan <segan@noreply.github.com> Cc: YinyanLiu <yinyan.yana.liu@gmail.com>, Mention <mention@noreply.github.com> Subject: Re: [santi-pdp/segan] SSNR is not as high as in the paper (#62) Hello, I have trained the SEGAN by myself and also use the pretrained model. I got the same metrics value of noisy as in the paper with SSNR=1.68 But when I evaluate the pretrained model, I only got SSNR=0.78, rather than 7.73 in the paper. I have try different methods to calculate the SSNR and all got the low SSNR. Anyone have the same issue? or have reappear the metrics in the paper? 你好，我想请教一下你的pesq，csig这些指标是怎么算的，可以指点一下嘛，谢谢你好，这些指标你计算了吗，我刚刚接触语音增强，指标还不太懂怎么计算 — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub, or unsubscribe. — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub, or unsubscribe.