论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning

虚幻大学 xuhss 523℃ 0评论

Python微信订餐小程序课程视频

https://edu.csdn.net/course/detail/36074

Python实战量化交易理财系统

https://edu.csdn.net/course/detail/35475阅读目录

论文地址:https://graz.pure.elsevier.com/en/publications/acoustic-echo-cancellation-with-cross-domain-learning

具有跨域学习的声学回声消除

回到顶部### 摘要:

  本文提出了跨域回声控制器(CDEC),提交给 Interspeech 2021 AEC-Challenge。该算法由三个构建块组成:(i) 时延补偿 (TDC) 模块,(ii) 基于频域块的声学回声消除器 (AEC),以及 (iii) 时域神经网络 (TD-NN)用作后处理器。我们的系统获得了 3.80 的整体 MOS 分数,而在 32 毫秒的系统延迟下仅使用了 210 万个参数。

关键字:声学回声消除、神经网络、残余回声消除

回到顶部### 1 引言

  回声消除 (AEC) 在当今的 VoIP 语音通信和视频会议系统中发挥着重要作用。由于室内声学,在扬声器和耳机麦克风、听筒或任何其他用于语音通信的音频硬件之间会出现回声。根据房间的混响时间,声学回声可能会非常突出,甚至会显着降低语音清晰度和语音质量 [1]。这在免提场景中尤其是一个问题 [2]。因此,高效的 AEC 解决方案是可靠语音通信的重要组成部分。典型的 AEC 将扬声器和麦克风之间的回声脉冲响应 (EIR) 建模为线性 FIR 滤波器,并使用归一化最小均方 (NLMS) 算法 [3, 4] 自适应地调整该滤波器。许多实现需要语音活动检测器 (VAD) 在双方通话期间停止适应,即当近端和远端说话者同时说话时 [3,5]。更复杂的实现通过使用状态空间模型 [6] 或卡尔曼滤波器 [7] 来解释双方对话。然而,线性回声模型不能考虑回声路径中的非线性失真,或麦克风拾取的附加噪声。 SpeexDSP [8]、WebRTC [9] 或 PjSIP [10] 等商业 AEC 框架依赖于传统的非线性回声和噪声去除方法,例如 Wienerfilters [11]、Volterra 内核 [12] 或 Hammerstein 模型 [13]。

  最近,已经提出神经网络用于非线性残余回声和噪声去除[14-19]。从深度学习的角度来看,这些任务可以看作是语音或音频源分离问题 [2,14,18-23]。尽管该研究领域近年来进展迅速 [24, 25],但大多数基于 NN 的说话人分离算法对计算的要求很高,没有因果关系,并且不能在实时应用中工作。能够进行实时处理的系统在逐帧的基础上运行。特别是,循环神经网络 (RNN),如门控循环单元 (GRU) [26] 或长短期记忆 (LSTM) [27] 网络用于模拟人类语音中的时间相关性,同时遵守实时典型 AEC 应用的约束 [2, 19, 28]。类似的架构 [29-31] 已应用于实时信号增强,作为对 Interspeech 2020 [32] 的深度噪声抑制挑战和 ICASSP AEC 挑战 [33] 的贡献。

  本文介绍了我们对 Interspeech 2021 AEC-Challenge 的贡献,该挑战由三个级联模块组成:(i) 基于 PHAse 变换的广义互相关 (GCCPHAT) [4] 的时延补偿 (TDC) 模块,其中补偿近端扬声器和麦克风信号之间的延迟。 (ii) 一种频域状态空间块分区 AEC 算法 [6],它去除了线性回波分量。 (iii) 时域神经网络 (TD-NN),它可以同时去除非线性残余回声和附加噪声。我们将我们的系统称为跨域回声控制器 (CDEC),因为它同时在频域和时域中运行。我们模型的评估基于使用 ITU P.808 框架 [33] 的感知语音质量指标,该框架报告平均意见分数 (MOS)。此外,我们报告了其他指标,例如 MOSnet [34] 和 ERLE [35]。最后,我们还报告了我们的 CDEC 系统在每帧音频数据的 MAC 操作方面的计算复杂性。

回到顶部### 2 提出的系统

2.1 问题表述

  在典型的 AEC 系统中,有两个输入信号可用: (i) 远端麦克风信号 x(t),由本地扬声器播放。 (ii) 近端麦克风信号 d(t),可描述为以下分量的叠加:

  bb4cf20e29ecf652259cf770a5a0731e - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning6c96866a65ac3b036de8c8e5c7afd425 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning5e3cc2b7732b19c035d3b78a33e9a80e - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learningd(t)=x(t−Δt)∗h(t)+s(t)+n(t)+v(t)  (1)
d(t)=x\left(t-\Delta_{t}\right) * h(t)+s(t)+n(t)+v(t)  (1)

其中 [外链图片转存失败,源站可能有防盗链机制,建议将图片保存下来直接上传(img-9LmljwlQ-1648529302429)(https://blog.csdn.net/2_7_5/fonts/HTML-CSS/TeX/png/Math/Italic/400/0064.png?V=2.7.5)][外链图片转存失败,源站可能有防盗链机制,建议将图片保存下来直接上传(img-scNiRxZR-1648529302430)(https://blog.csdn.net/2_7_5/fonts/HTML-CSS/TeX/png/Main/Regular/400/0394.png?V=2.7.5)]![](https://img-blog.csdnimg.cn/img_convert/d28facf247930cdcdbd0a7a70c772c6f.png)Δt\Delta\_{t}建模为滤波器权重中的前导零。然而,更实际的做法是在 AEC 之前明确补偿此延迟,以保持建模的 EIR 较短,从而节省计算资源。虽然这种延迟可能是未知的,但我们坚持通常在实时音频处理框架中做出的假设 [8-10]。特别是,我们假设延迟不超过1s,并且至少稳定10s。

  我们采用 GCC-PHAT 算法 [4] 在频域中比较远端信号25c9785990c180870e7a26b78296d94b - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learningd5fe6ee39c9cc36a7ebe9b0f6e4250d0 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning29a6b2e6bc9fb8b4f3d038850b745126 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning71d75bb6e6f0922707929ee616326b19 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learningcf7cf93547566fd30e448fa80602e0d4 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning08c3cb30299ab4210945da1af4f92ab3 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learningdfb9976f153cb198e9f963ce32319e3b - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning30fc45e261ef23ef99ae4176d0e22785 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learningd5fe6ee39c9cc36a7ebe9b0f6e4250d0 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning29a6b2e6bc9fb8b4f3d038850b745126 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning71d75bb6e6f0922707929ee616326b19 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning0fe75d486e243e09ecd0ba6ae7988245 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning095d20e0673beb4abfa546adce933b8d - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learninge73a2670b7c5e4bfe2e992b5a4bc599a - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning2fb3b3449e595bfc6dba0398284bed4a - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learningbe956450e539bc5c88d2193e7e2a42c9 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning015dba733b98e28073acb8b488dfe27a - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning095d20e0673beb4abfa546adce933b8d - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning62dbec774af6735c30a79374a3951008 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning6fae1b968be93bec0e63af20edf0311a - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learningca1d33c636add1412d989b312d73b348 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learningd5fe6ee39c9cc36a7ebe9b0f6e4250d0 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning29a6b2e6bc9fb8b4f3d038850b745126 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning71d75bb6e6f0922707929ee616326b19 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning227d00a4ec54135d3c1a3c748d786825 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning45ef6d38166a136b6b7679e56e6855cd - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning619712610d2ceb925372aaa4ac9a68e8 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learningd5fe6ee39c9cc36a7ebe9b0f6e4250d0 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning29a6b2e6bc9fb8b4f3d038850b745126 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning71d75bb6e6f0922707929ee616326b19 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learningb8ac40b8f83205725e32b77ec4234c99 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning879e24e93fb2a8926cd258ba718384dd - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning  cfb6fa615ed52fd8ad38064856cd36fd - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learningd380f5e23c0ff79488e3116ff1e74c80 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learninga8de04563010fb846095b693cc35529a - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain LearningΦ(l,k)=Φ(l,k)α+(1−α)X(l,k)D(l,k)∗  (2)

\Phi(l, k)=\Phi(l, k) \alpha+(1-\alpha) X(l, k) D(l, k)^{*}  (2)

其中5df2c5a61c036454da1af7441b8d689e - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning22c172fc3236bd6a644360f3565509c2 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning34fd99c434e0ff2a21cc211943f9c477 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning061be77b4d5baf27ff92e5d9381682f1 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning0adb6a7ca141aca2c7546af1c1c20529 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning7627ca6a693fc2f4c934d4e3bbbce69e - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning34fd99c434e0ff2a21cc211943f9c477 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learningbaf1be5dc9fbeafada54b64d78ea9ac6 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learningfe04475663746de7caaf02c31eb30d9c - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning19ff3b79ce17c23a25c0884f315e7103 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learningaa665bbb834f7de1f24f65eda076a8ba - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning87ab8be46a87b3b19b123e35c5bf11cc - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learningdfb9976f153cb198e9f963ce32319e3b - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learningd6859a5e5ccadb3106648005c7a67f54 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learningd5fe6ee39c9cc36a7ebe9b0f6e4250d0 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning8641f0beaa5b90d14cfe4fb5b7d0c431 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning2e22fa5a5587b5119ea4f705cdef3525 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learningdfb9976f153cb198e9f963ce32319e3b - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learningadff903044f4088ddff50a97aa3bdbc3 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learningd5fe6ee39c9cc36a7ebe9b0f6e4250d0 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning68ab49b6ff412c2417f58a4fa5d3ef84 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learningf28998bde24aa0b9458e7220602f8b6f - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning  a6e54f52d39e08b0d91272c8b37e097a - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learningc274cbcddff695af95eb4ac112bf624f - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learningdce0605d4d393a8cd492bf4ab9ec1bba - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain LearningΔt=argmaxtF−1Φ(l)|Φ(l)|  (3)
\Delta_{t}=\underset{t}{\operatorname{argmax}} \mathcal{F}^{-1} \frac{\Phi(l)}{|\Phi(l)|}  (3)

其中[外链图片转存失败,源站可能有防盗链机制,建议将图片保存下来直接上传(img-cZyu02RY-1648529302476)(https://blog.csdn.net/2_7_5/fonts/HTML-CSS/TeX/png/Math/Italic/400/0064.png?V=2.7.5)][外链图片转存失败,源站可能有防盗链机制,建议将图片保存下来直接上传(img-5ZcUciz2-1648529302477)(https://blog.csdn.net/2_7_5/fonts/HTML-CSS/TeX/png/Math/BoldItalic/400/0078.png?V=2.7.5)][外链图片转存失败,源站可能有防盗链机制,建议将图片保存下来直接上传(img-MPSnWeb9-1648529302477)(https://blog.csdn.net/2_7_5/fonts/HTML-CSS/TeX/png/Math/Italic/400/0064.png?V=2.7.5)][外链图片转存失败,源站可能有防盗链机制,建议将图片保存下来直接上传(img-eZJGzSdO-1648529302478)(https://blog.csdn.net/2_7_5/fonts/HTML-CSS/TeX/png/Main/Regular/400/0032.png?V=2.7.5)]![](https://img-blog.csdnimg.cn/img_convert/850fa531b013cb09e96e3ecabff2fe4e.png)2T2T 最新样本,即

  e61368ccce6a70b7418cb110f125e904 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learningba108dd1626e35931b9351059f0d0791 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning5132fecc8e75d7f0a7e5d073c633acb9 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learningx′(l)=x(t+n−2T)d′(l)=d(t+n−2T)  (4)
\begin{aligned}

&\boldsymbol{x}^{\prime}(l)=x(t+n-2 T) \

&\boldsymbol{d}^{\prime}(l)=d(t+n-2 T)  (4)

\end{aligned}

其中PP 个分区,即

  a30c4d838b9c7b4162e55ef5570346e5 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning474a79e0e8424f1943762e6f54c5aef2 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learninga5c74ad47fdad1e9cf6f0be51d88486a - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain LearningY(l,k)=∑p=0P−1X(l−p,k)W(p,k)e′(l)=d′(l)−F−1{Y(l)}  (5)
\begin{aligned}

Y(l, k) &=\sum_{p=0}^{P-1} X(l-p, k) W(p, k) \

e^{\prime}(l) &=\boldsymbol{d}^{\prime}(l)-\mathcal{F}^{-1}{\boldsymbol{Y}(l)}

\end{aligned}  (5)

其中379ce1f86b83611cc6a914006904fb6f - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning0fb16543e0a37f1508ea6ba8f3276d10 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learningf396cfc822499a1965882ac5a31ea073 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learningc88eaee508d3eada52f2ad050c06ab2d - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learningd9cfb982f60caa6d8210ca4aa401741c - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learningd81875b9de94d04aec450a63894201e9 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learninga5c74ad47fdad1e9cf6f0be51d88486a - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning6065d43d864e5b48b7e26a9ac8dcea06 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning8686fa9c84a071cf139ad1e88cc1cfa4 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learningc9c5ba1d9f00337a90679d58ae66ce00 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learninge53c331d9d26e33c78d136af14d1e66f - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learningd5fe6ee39c9cc36a7ebe9b0f6e4250d0 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning6577e912029f0b39a8a452c97ae78c59 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning668394ab4eeca7d2e4761acb8db13528 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learninge1e98c617b7871162d94b3e7da3cabb9 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learninga0a8854ebe5e08321f213908b75d17cf - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning6d6a300ce1635e7a06abe9ebb371d899 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning  236fc41eb4919fa767e9b658c388168b - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning520a34a7422e0b027265d4a352c314ae - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learningf4c6bff037dc5ed56e38cdd7cf5550eb - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learninge(t+n−T)=e′(l,n+T)  (6)
e(t+n-T)=e^{\prime}(l, n+T)  (6)

其中,a773c0dbf710b2c6ac18199b3bcc9426 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning6ae0e02f0f50a6bb7a72ebbf08919a51 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning6c56dd044a3eb61bfffd74a6c51e5e78 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning864b050c185a1dff00cf0574e8a2b972 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning30ad09be33878a2e50696132fb4a351e - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning6ae0e02f0f50a6bb7a72ebbf08919a51 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning07ddccd2a842b3687a71bfebd1b95842 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning36b612a29543d1b1a49a6776f093c545 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning20945d5d4f655ec743cf9599d8006eb2 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learningecc344db453ec489cc76b2c88d52ae0f - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learninga7ec74c6e3ccd1f6193d6b7e1b650de4 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning3f4b920705a898efd563c3ad17f8bc5b - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning8e4ba056b8a7748ab29792d182b34780 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning6ae0e02f0f50a6bb7a72ebbf08919a51 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learningebef1fbb8cfbf63961add9cca4e1bc00 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learningaf275106ae1b563aaa1a5a281dac7230 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning19ff3b79ce17c23a25c0884f315e7103 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning119676b06a675598a9d4c479c80ecd3d - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning58f9dfaeb341926070a107b7a539a56c - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning6732f75d3842b4df1c5c9772d92e460a - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning3f4b920705a898efd563c3ad17f8bc5b - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learningd66a72adff5de6a964981826d9fd4d73 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning6ae0e02f0f50a6bb7a72ebbf08919a51 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learningc79bd92ac42e2c278f7f9d7dcaa1e065 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learningb7df7bc5226a1eb0bf9ad881446aaa2c - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learningba0c65891793e21d88e25aa5d0f69237 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learningaa594020b0c022f1c98e08d0efcb67bf - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learningeaeaebf328be66892b0ce22b5033474d - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning5eaee33726545be3be96f04e6e4600de - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning19ff3b79ce17c23a25c0884f315e7103 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning29cefe1daabc4ccf476c98fc6a0fe95b - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learningdd6e881ab748fb0445ef4f989fc4b8bb - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning3d540dfa26b7252601b885fd7e683b9d - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learningde33aa12266d811d646820f388e2850d - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning165c7a6d2a33fff57d49250bce726c46 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning1a5634037c6f840db3f870bfbed303b2 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning  f0f816bd798c51102602f10f6031c5c5 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning2ce7f47387ee95833db2c1745a443590 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning499f12429c12d3de9efff2331086d939 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learningw(p)=F−1{W(p)},w(p,n+T)=0W(p)←F{w(p)}  (7)
\begin{aligned}

\boldsymbol{w}(p) &=\mathcal{F}^{-1}{\boldsymbol{W}(p)}, \

w(p, n+T) &=0 \

\boldsymbol{W}(p) & \leftarrow \mathcal{F}{\boldsymbol{w}(p)}

\end{aligned}  (7)

其中,[外链图片转存失败,源站可能有防盗链机制,建议将图片保存下来直接上传(img-M60FFWr3-1648529302525)(https://blog.csdn.net/2_7_5/fonts/HTML-CSS/TeX/png/Math/Italic/400/0057.png?V=2.7.5)]![](https://img-blog.csdnimg.cn/img_convert/565171cdc9b326067de2297c5118ccd7.png)![](https://img-blog.csdnimg.cn/img_convert/0699c8bf2a37cd9e1ce32d98eee30f26.png)![](https://img-blog.csdnimg.cn/img_convert/58dc5acd3f68f22db8caaf08efeb899f.png)![](https://img-blog.csdnimg.cn/img_convert/bac94994600c352292f8d6d3eae89879.png)![](https://img-blog.csdnimg.cn/img_convert/d92409e0e1dc9c6b63cffbaf0035e388.png)![](https://img-blog.csdnimg.cn/img_convert/43c0d3f660d747ff142a5ae97ee065f6.png)W^(p,k)\hat{W}(p, k)作为阴影权重。算法 1 说明了这些权重是如何更新的。

d3c37530c0c802c569a7a32bbc7b1fec - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning

阴影权重基于 ERLE 69322df1844b33c23ffe9aee40bd712c - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning8d36a599e5f0535f6b975bab7bba7cf6 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning499f12429c12d3de9efff2331086d939 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learningea5c52c89ffaf3c2b1a3b0fa6a587d97 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learningd16a48bc59ab72e7b5364c477a394dfc - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learningb3a0d42120b88d474181c470bb44fc0c - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning471f70557cfe3641578f7c92da22b784 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning2e72efdff4a4a0c50913a9495b65b01e - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning4150561999581f28691e1420f81f5d5d - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learninga0b42cf8021925a3738c104e6d8dbcad - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learningee6c5f591b2d16f1291c2128dd42f26b - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning9e109851ed67645c6b5504042f5fa388 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning4153519d511ddfdab0e3e8343f427eb1 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learningabcd286a800e7a746fa9d2d6541cea5c - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learningf4552e727deab5a76231d16d8fdf6b22 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning393e917065755a608204032c40f654e2 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learningbc3c04f2f186533a5d23e8f2ff9cccee - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning331f8e13633b3e317065ffcfaa66ee1b - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learningda02595fd5a1eb619d065965dec93071 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learningea69021409dd0f5120bf42ec7fc4cba6 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learninge61cb9878fad7d787d6c64b618e584b7 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learningd656f2b226a661448d1990fc7486fcdb - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learninge480b747f878cdffe3219fbe6742c14e - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning005aaec39a2224fe664987950cc9485f - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning7f7817ad9c31382ec46de2b2d48c5bed - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learningc2a0f2498751d2b9e1ead502e7708e28 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learningbcc9f48935740249902bf1ac92471ebb - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning3ebde291ca2bb6c47139a6f488540b3b - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learningafdfbb080da74b0578bb9b927d269a2e - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning353b479953424e62bbce19d351085c5e - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning8bfa26f6dbf8a7fef9612b49a3ff42d0 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning1dca70e9ee92414127c3b044625ef45f - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning14b2fcbcb7b4fd535e77e9b1cd3692e4 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learningb5acfe81badfc4119b97eb3a3fd92c3a - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning565171cdc9b326067de2297c5118ccd7 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning40ddf7ed3aa84046a4cace460b5c8883 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learningb1e0f692d1353f950243f6d0c3fc2f44 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning4785b08a5262930ccfefc02aef440737 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning33e520089182d596257bbe7e34cee347 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learningfe85c42cfd5225b84725e5f4841a03ca - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learningca7358edd6e67602edc741ad9522ad71 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learningb3afa1d34c0f797518b095fd70d90dc4 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning2e72efdff4a4a0c50913a9495b65b01e - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning049de355e525a0c1bb032343927ff844 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning6a5a3037d61bd2a0eb1cc3c8d60bc435 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learningee6c5f591b2d16f1291c2128dd42f26b - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning9e109851ed67645c6b5504042f5fa388 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning005aaec39a2224fe664987950cc9485f - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learningd71c2e76ea3e05cdc35ea1ab1e195c09 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learningb0b53755bbebf054a694880806f2f749 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning11ef00bf453ec9bb12b77dd776c8dedf - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learningabfecc45e007c1b05c095d51a1a0ebd6 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning3610ca02eedb54819e2cc8050fad64d4 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learningcf135b34ab467db94e39619819602347 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning10505ef8eae2b3bddfb96bfc935055c8 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning60795e28253a31d75e1bba91aa59950d - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning88b5d84349640bf8eb046912f547cf89 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning311334d2e63f2db57a08cfab403f99cd - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning9409e998592190dd129fc0357b32b99f - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learningebbd0e2f81db7b49f15e1ff290720922 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning4a06f4b0fa6b2def5a4077599bb90536 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning565171cdc9b326067de2297c5118ccd7 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learningc14a982636431df0c67f1c4f683b5c4d - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning666088829f342b935555cbe0359e7445 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learningc5d7e835b57ad17a07807573e0d7630b - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning1006952ac6aa99c83d424ebd37fcfbe0 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning1162401b13247e64079a48ff3fcfa2d8 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learninga7e20b281f9d14dc7d63bb1902f81de0 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learningd633f2876737a91aaee894f4286c9e27 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning  5992359f208cd11c4b6fce3f5ccb4c43 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learningafd6dd0d48b848747c37f9aa734720ce - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning7defc7fd20b7eef568abee3f5a26a858 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain LearningE(l)=10log10⁡∑k|D(l,k)|2∑k|E(l,k)|2E^(l)=10log10⁡∑k|D(l,k)|2∑k|E^(l,k)|2  (8)
\begin{gathered}

\mathcal{E}(l)=10 \log _{10} \frac{\sum_{k}|D(l, k)|^{2}}{\sum_{k}|E(l, k)|^{2}} \

\hat{\mathcal{E}}(l)=10 \log _{10} \frac{\sum_{k}|D(l, k)|^{2}}{\sum_{k}|\hat{E}(l, k)|^{2}}

\end{gathered}  (8)

其中[外链图片转存失败,源站可能有防盗链机制,建议将图片保存下来直接上传(img-nNYTElau-1648529302585)(https://blog.csdn.net/2_7_5/fonts/HTML-CSS/TeX/png/Math/Italic/400/0046.png?V=2.7.5)][外链图片转存失败,源站可能有防盗链机制,建议将图片保存下来直接上传(img-Ks5h7mAd-1648529302585)(https://blog.csdn.net/2_7_5/fonts/HTML-CSS/TeX/png/Math/Italic/400/0053.png?V=2.7.5)][外链图片转存失败,源站可能有防盗链机制,建议将图片保存下来直接上传(img-cXphPRxl-1648529302586)(https://blog.csdn.net/2_7_5/fonts/HTML-CSS/TeX/png/Math/Italic/400/0078.png?V=2.7.5)][外链图片转存失败,源站可能有防盗链机制,建议将图片保存下来直接上传(img-3N6PMd23-1648529302587)(https://blog.csdn.net/2_7_5/fonts/HTML-CSS/TeX/png/Math/Italic/400/0079.png?V=2.7.5)][外链图片转存失败,源站可能有防盗链机制,建议将图片保存下来直接上传(img-avInbcRx-1648529302588)(https://blog.csdn.net/2_7_5/fonts/HTML-CSS/TeX/png/Math/Italic/400/0064.png?V=2.7.5)][外链图片转存失败,源站可能有防盗链机制,建议将图片保存下来直接上传(img-der1EsSX-1648529302589)(https://blog.csdn.net/2_7_5/fonts/HTML-CSS/TeX/png/Math/Italic/400/0065.png?V=2.7.5)][外链图片转存失败,源站可能有防盗链机制,建议将图片保存下来直接上传(img-yXDhehP6-1648529302589)(https://blog.csdn.net/2_7_5/fonts/HTML-CSS/TeX/png/Math/Italic/400/0048.png?V=2.7.5)]HH 个神经元的潜在表示。请注意,此 Conv1D 层使用过去 1600 个相应信号的样本,即它看到过去 100 毫秒音频数据的上下文。每个信号都通过即时层归一化单独归一化,以解决各个级别的变化。即时层归一化类似于标准层归一化[36]。该分支中的最后一个前馈 (FF) 层使用 softplus 激活函数,以提供不受约束的掩码。

ab411771837f5fa1fca2050d60f772fd - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning

  图 2 中的下部分支说明了将掩码应用于潜在空间中的残差信号 [外链图片转存失败,源站可能有防盗链机制,建议将图片保存下来直接上传(img-HU8B2JsQ-1648529302591)(https://blog.csdn.net/2_7_5/fonts/HTML-CSS/TeX/png/Math/Italic/400/0064.png?V=2.7.5)][外链图片转存失败,源站可能有防盗链机制,建议将图片保存下来直接上传(img-bGTmDCqL-1648529302592)(https://blog.csdn.net/2_7_5/fonts/HTML-CSS/TeX/png/Math/Italic/400/0064.png?V=2.7.5)]![](https://img-blog.csdnimg.cn/img_convert/fedcacede8161cac3d58df2f8c35571f.png)![](https://img-blog.csdnimg.cn/img_convert/943ceb01c87230f132f635a2e41b2213.png)![](https://img-blog.csdnimg.cn/img_convert/376b5de5763ea03403a334a2d97f0992.png)![](https://img-blog.csdnimg.cn/img_convert/104cfc98958be149d230bfc9fc3ab9c7.png)![](https://img-blog.csdnimg.cn/img_convert/b45c76e3f80b919b00b745bbabc065f6.png)![](https://img-blog.csdnimg.cn/img_convert/75dc13da3ce634ff745bcac525ae00ed.png)![](https://img-blog.csdnimg.cn/img_convert/3f3113539b91724443ce381c1f6c100c.png)![](https://img-blog.csdnimg.cn/img_convert/df28c4f0a57df59a74045660aaaa11eb.png)![](https://img-blog.csdnimg.cn/img_convert/70758ba3f29a465db8768d3b6b3a511d.png)![](https://img-blog.csdnimg.cn/img_convert/a690b2e1b88509eb64b5e4e324ff38da.png)![](https://img-blog.csdnimg.cn/img_convert/84c51dd81ed367dae59e25071902f176.png)d(t)>−40dBFSd(t)>-40 d B_{F S}的平均能量。否则信号被拒绝。

  我们通过将干净的 WSJ0 数据 [37] 作为所需的近端语音 [外链图片转存失败,源站可能有防盗链机制,建议将图片保存下来直接上传(img-qLAHoMQh-1648529302601)(https://blog.csdn.net/2_7_5/fonts/HTML-CSS/TeX/png/Main/Regular/400/0031.png?V=2.7.5)]![](https://img-blog.csdnimg.cn/img_convert/5f7d032513e3cce796786b7decbdfe62.png)![](https://img-blog.csdnimg.cn/img_convert/e9e8ecd70e97e450c6ffa05ae552eaef.png)![](https://img-blog.csdnimg.cn/img_convert/827dd23bd8c9abe5fadc80f3e4dfe8f4.png)![](https://img-blog.csdnimg.cn/img_convert/3418d26209edbb85e771f9ffb28c031d.png)![](https://img-blog.csdnimg.cn/img_convert/df28c4f0a57df59a74045660aaaa11eb.png)![](https://img-blog.csdnimg.cn/img_convert/16b644cb2cbf0bd820ea79a6f9addda0.png)12…36dB12 \ldots 36 dB之间的均匀分布中随机选择。噪声仅添加到模拟数据集中。

  为了进一步提高鲁棒性并模拟各种传输效果,我们在模拟数据集的每个远端信号 [外链图片转存失败,源站可能有防盗链机制,建议将图片保存下来直接上传(img-JJoflkII-1648529302605)(https://blog.csdn.net/2_7_5/fonts/HTML-CSS/TeX/png/Math/Italic/400/0064.png?V=2.7.5)]![](https://img-blog.csdnimg.cn/img_convert/f9dca30ee1d61a1ba028637eddb8c099.png)![](https://img-blog.csdnimg.cn/img_convert/e67b7d93bf6108151f3302564e7bedf5.png)![](https://img-blog.csdnimg.cn/img_convert/b2a9f6b9e1b3c8035e55b4ffaf9351a7.png)d(t)d(t)。我们将每个信号截断为 10 秒,以便能够将它们堆叠成批次进行训练。

3.3 CDEC训练

  在训练期间,我们首先使用式2中的 GCCPHAT 估计批量延迟。 每 10 秒一次,即每个训练话语一次。接下来,我们从式5-8执行 AEC。输出回波模型 f4d31de6e1ea541ddc35d5f68e92c0ba - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning0cb03ad202fa3289cbb7be965f16e12c - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning64a901e51e10e8d8aaefe42fb1e687a0 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning3adc671506034cda86e8f73e7ecd3239 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learningeba39053a0cb555316669d29e5a6a0ec - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning3092842c21d81c0273c48bb5cdc325dd - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning471f70557cfe3641578f7c92da22b784 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning8d158b809b456429dee5067fc3180d9d - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learninga7e26eeaf6fb8333de46f21c3c200510 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learningd0190f9133a696a3659e7c985ea5e54f - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learningee6c5f591b2d16f1291c2128dd42f26b - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning9e109851ed67645c6b5504042f5fa388 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning635497dacca8caf22aa0e03d821b53a1 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning34c16d6d985085614f638b541e011ded - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learningb2878ac419a2965d5c78af1c214198f8 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learningb226f96a7542857eebeb5315dc135f0d - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning7abbf40cb9cf2cf6a3545398ecc7f203 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning7b69e86fdb807e4d669afddb22b71c32 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning9e109851ed67645c6b5504042f5fa388 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning2b2bf4b4f284430979ace6d1917b0d9e - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning8b973b50cc4ab0d7b8b02701f1865e16 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning34c16d6d985085614f638b541e011ded - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning704338b0177f7985a206254177bdd278 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learningbd80263a6ea97456ce5760ac2b8340a5 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning3454b67f5d462805ffaa86097aecce2c - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning9320528e222b4fa9f356aadb3e16d681 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learninge1af86a61e52d80a288d247cf277d934 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learningdd807a27c1edb83f530f89a8cf2738cb - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning0d3ecabc089af1987600e85358579410 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning15f8ab036ccc73e6964ce07ae3e417ba - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning380a0120f214951879a4acbc30b711c6 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning915b2532896510d0e6374d350434a3fd - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning  c01939fbbece88a7fc1288a09b251ad1 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learningac90b5efc6821304a388cbc98a6b298c - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning2c6e888e56caaf26547515aa89b2601e - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain LearningLSDR=10log10⁡∑ts(t)2∑t[s(t)−z(t)]2  (9)
\mathcal{L}_{\mathrm{SDR}}=10 \log _{10} \frac{\sum_{t} s(t)^{2}}{\sum_{t}[s(t)-z(t)]^{2}}  (9)

  而我们使用 ERLE 作为 FE 场景的损失函数,即

  7bb394e1b036bb991af328941d6da65f - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learninge5aa37c2425c84a86b9395a8e0ae40fa - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning0ce1a6bff274ab776565e140e19f6cc2 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning07424f66a8adfb2175ead37521dcd8be - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain LearningLERLE=10log10⁡∑td(t)2∑tz(t)2  (10)
\mathcal{L}_{\mathrm{ERLE}}=10 \log _{10} \frac{\sum_{t} d(t)^{2}}{\sum_{t} z(t)^{2}}  (10)

  我们将总损失函数定义为

  237e58ac2858a1d7623a956ce87988b8 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learningefffd9d56bf0a63d579fa9fe6fb21bc4 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning7ac0969d23037392f1f09be5db176ae1 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learningfff4bc231172c647078b0b961dc24789 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain LearningLERLE=−LSDR−λLERLE  (11)
\mathcal{L}_{\mathrm{ERLE}}=-\mathcal{L}_{\mathrm{SDR}}-\lambda \mathcal{L}_{\mathrm{ERLE}}  (11)

我们设置[外链图片转存失败,源站可能有防盗链机制,建议将图片保存下来直接上传(img-sILakpD7-1648529302641)(https://blog.csdn.net/2_7_5/fonts/HTML-CSS/TeX/png/Main/Regular/400/0394.png?V=2.7.5)][外链图片转存失败,源站可能有防盗链机制,建议将图片保存下来直接上传(img-cROWYi5j-1648529302641)(https://blog.csdn.net/2_7_5/fonts/HTML-CSS/TeX/png/Math/Italic/400/0066.png?V=2.7.5)][外链图片转存失败,源站可能有防盗链机制,建议将图片保存下来直接上传(img-r9NecrF5-1648529302642)(https://blog.csdn.net/2_7_5/fonts/HTML-CSS/TeX/png/Main/Regular/400/0032.png?V=2.7.5)][外链图片转存失败,源站可能有防盗链机制,建议将图片保存下来直接上传(img-ZPcb1bqE-1648529302643)(https://blog.csdn.net/2_7_5/fonts/HTML-CSS/TeX/png/Math/Italic/400/0078.png?V=2.7.5)][外链图片转存失败,源站可能有防盗链机制,建议将图片保存下来直接上传(img-8ru1xYVp-1648529302644)(https://blog.csdn.net/2_7_5/fonts/HTML-CSS/TeX/png/Math/Italic/400/0064.png?V=2.7.5)][外链图片转存失败,源站可能有防盗链机制,建议将图片保存下来直接上传(img-a6u6gvLJ-1648529302644)(https://blog.csdn.net/2_7_5/fonts/HTML-CSS/TeX/png/Math/Italic/400/0065.png?V=2.7.5)][外链图片转存失败,源站可能有防盗链机制,建议将图片保存下来直接上传(img-cuSGNaCU-1648529302645)(https://blog.csdn.net/2_7_5/fonts/HTML-CSS/TeX/png/Math/Italic/400/0046.png?V=2.7.5)][外链图片转存失败,源站可能有防盗链机制,建议将图片保存下来直接上传(img-N6UHthbg-1648529302646)(https://blog.csdn.net/2_7_5/fonts/HTML-CSS/TeX/png/Math/Italic/400/0053.png?V=2.7.5)][外链图片转存失败,源站可能有防盗链机制,建议将图片保存下来直接上传(img-AacE4I3r-1648529302646)(https://blog.csdn.net/2_7_5/fonts/HTML-CSS/TeX/png/Math/Italic/400/007A.png?V=2.7.5)][外链图片转存失败,源站可能有防盗链机制,建议将图片保存下来直接上传(img-PqT10wik-1648529302649)(https://blog.csdn.net/2_7_5/fonts/HTML-CSS/TeX/png/Main/Regular/400/0032.png?V=2.7.5)]![](https://img-blog.csdnimg.cn/img_convert/3d5c493bf177807d183025644c62f492.png)2T2T 个样本,相当于 32ms。由于所有三个模块都在相同的块上运行,因此 CDEC 系统的总延迟为 32 毫秒。

4.3 计算复杂度

  CDEC 模型的计算复杂度在四核 i5 2.5Ghz 参考系统上进行了评估。特别是,我们使用矩阵/向量库 Eigen 和 FFT 库 FFTW [39] 使用 C++ 中的单精度参考实现测量了 CDEC 系统前向传递的单帧的执行时间。 TD-NN 系统使用 210 万个参数,而 AEC 使用<span class="MathJax" id="MathJax-Element-127-Frame" tabindex="0" style="position: relative;" data-mathml='

8593f1ea230b0cb9e8aa8b5a993964fc - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning26f36f49fbb8c89e9fb3f709feb6f282 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning2af2023acbe54e767551a806a82fda3a - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning868589a42088a84afd4c042377567ccd - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning13c35479814c8dd6cccf86ef06deeda2 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning4df7a3b4af99bf159bb5ecbf35f0759c - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning2237d0d7b14d26933268cd8fa25fa018 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning243aff9cb984751daa079b10d8efe8c6 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learninga005a5c2d7485dca38cfacc6eedc0679 - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learningc5a0aca49ef63cb8eb1967e77e57a7dc - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning2P⋅2T=163842 P \cdot 2 T=16384个复值滤波器权重,包括阴影权重。一个推理步骤每帧需要 228 us。特别是,TDC 模块需要 0.16 us,AEC 32.88 us 和 DNN 195 us 处理单帧。请注意,TDC 模块每 10 秒执行一次,因此它对单帧执行时间的贡献相当小。 CDEC 系统的整体执行时间是使用单个 CPU 时每 1 秒音频的 28.8 毫秒。在 XNNPACK的帮助下进行多线程执行的情况下,运行时间可以减少到 19.6 毫秒,处理一秒的音频。该模型可以使用稀疏格式和剪枝进一步缩小。此外,定点表示的使用大大降低了内存消耗和计算复杂度,如 [40] 所示。表 3 显示了 CDEC 模型的计算复杂度。

表 3:单精度 CDEC 模型的计算复杂度,在四核 i5 2.5GHz 参考系统上测量。

bfe5e2d989798eb499114eea861fd74f - 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning

回到顶部### 5 结论

  我们展示了我们的跨域回声控制器 (CDEC)——一种为 2021 年语音间 AEC 挑战赛开发的实时 AEC 系统。该系统由三个模块组成,即延时补偿 (TDC) 模块、基于频域模块的声学回声消除器 ( AEC) 和时域神经网络 (TD-NN)。使用 ITU P.808 众包框架对 CDEC 进行了单声道和双声道回声场景的评估。特别是,它使用具有 2.1M 参数的模型报告了 3.80 的平均 MOS 分数。整个系统的整体延迟为 32ms,在 2.5 Ghz 四核 i5 系统上实时系数为 0.0288。

回到顶部### 6 参考文献

[1] H. Kuttruff, Room Acoustics, 5th ed. London–New York: Spoon Press, 2009.
[2] L. Pfeifenberger and F. Pernkopf,“Nonlinear Residual Echo Suppression Using a Recurrent Neural Network,” in Proc. Interspeech 2020, 2020, pp. 3950–3954.
[3] S. Haykin, Adaptive Filter Theory, 4th ed. New Jersey: Prentice Hall, 2002.
[4] J. Benesty, M. M. Sondhi, and Y. Huang, Springer Handbook of Speech Processing. Berlin–Heidelberg–New York: Springer,2008.
[5] G. Enzner, H. Buchner, A. Favrot, and F. Kuech,“Chapter 30 acoustic echo control,” in Academic Press Library in Signal Processing: Volume 4, ser. Academic Press Library in Signal Processing, J. Trussell, A. Srivastava, A. K. Roy-Chowdhury, A. Srivastava, P. A. Naylor, R. Chellappa, and S. Theodoridis, Eds. Elsevier, 2014, vol. 4, pp. 807–877.
[6] F. Kuech, E. Mabande, and G. Enzner,“State-space architecture of the partitioned-block-based acoustic echo controller,” in IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2014, pp. 1295–1299.
[7] C. Wu, X. Wang, Y. Guo, Q. Fu, and Y. Yan,“Robust uncertainty control of the simplified kalmanfilter for acoustic echo cancelation,” Circuits, Systems, and Signal Processing, vol. 35, no. 12,pp. 4584–4595, 2016.
[8] “Speex-dsp,” Website, visited on February 19th 2020. [Online].Available: https://github.com/xiongyihui/speexdsp-python
[9] “Webrtc toolkit,” Website, 2011, visited on March 14th 2021.[Online]. Available: https://webrtc.org/
[10] “PjMedia– acoustic echo cancellation api.” Website, 2008, visited on March 13th 2021. [Online]. Available: https://www.pjsip.org/pjmedia/docs/html/group PJMEDIA Echo Cancel.htm
[11] H. Huang, C. Hofmann, W. Kellermann, J. Chen, and J. Benesty,“A multiframe parametric wienerfilter for acoustic echo suppression,” in IEEE International Workshop on Acoustic Signal Enhancement (IWAENC), 2016, pp. 1–5.
[12] F. Kuech and W. Kellermann,“A novel multidelay adaptive algorithm for volterrafilters in diagonal coordinate representation [nonlinear acoustic echo cancellation example],” in IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP), vol. 2, 2004, pp. ii–869.
[13] S. Malik and G. Enzner,“Fourier expansion of hammerstein models for nonlinear acoustic system identification,” in IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2011, pp. 85–88.
[14] C. M. Lee, J. W. Shin, and N. S. Kim,“Dnn-based residual echo suppression,” in Interspeech, 2015.
[15] T. V. Huynh,“A new method for a nonlinear acoustic echo cancellation system,” in International Research Journal of Engineering and Technology, vol. 4, 2017.
[16] H. Zhang and D. Wang,“Deep learning for acoustic echo cancellation in noisy and double-talk scenarios,” in Interspeech, 2018,pp. 3239–3243.
[17] G. Carbajal, R. Serizel, E. Vincent, and E. Humbert,“Multiple-input neural network-based residual echo suppression,” in IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2018, pp. 231–235.
[18] G. Carbajal, R. Serizel, E. Vincent, and E. Humbert,“Joint dnn based multichannel reduction of acoustic echo, reverberation and noise,” CoRR, vol. abs/1911.08934, 2019.
[19] Q. Lei, H. Chen, J. Hou, L. Chen, and L. Dai,“Deep neural network based regression approach for acoustic echo cancellation,”in International Conference on Multimedia Systems and Signal Processing (ICMSSP). New York, NY, USA: Association for Computing Machinery, 2019, p. 94–98.
[20] L. Pfeifenberger, M. Zöhrer, and F. Pernkopf,“Eigenvector-based speech mask estimation for multi-channel speech enhancement,”IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 27, no. 12, pp. 2162–2172, 2019.
[21] L. Pfeifenberger, M. Zöhrer, and F. Pernkopf,“Dnn-based speech mask estimation for eigenvector beamforming,” in 2017 IEEE International Conference on Acoustics, Speech and Signal Processing, ICASSP 2017, Mar. 2017, pp. 66–70.
[22] H. Zhang, K. Tan, and D. Wang,“Deep learning for joint acoustic echo and noise cancellation with nonlinear distortions,” Interspeech, pp. 4255–4259, 2019.
[23] A. Fazel, M. El-Khamy, and J. Lee,“CAD-AEC: context-aware deep acoustic echo cancellation,” in IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP).IEEE, 2020, pp. 6919–6923.
[24] J. R. Hershey, Z. Chen, J. Le Roux, and S. Watanabe,“Deep clustering: Discriminative embeddings for segmentation and separation,” in IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2016, pp. 31–35.
[25] M. Kolbak, D. Yu, Z. Tan, and J. Jensen,“Multitalker speech separation with utterance-level permutation invariant training of deep recurrent neural networks,” IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 25, no. 10, pp. 1901–1913, 2017.
[26] J. Chung, C¸ . Gülc¸ehre, K. Cho, and Y. Bengio,“Empirical evaluation of gated recurrent neural networks on sequence modeling,”CoRR, vol. abs/1412.3555, 2014.
[27] S. Hochreiter and J. Schmidhuber,“Long short-term memory,”Neural Computation, vol. 9, no. 8, p. 1735–1780, 1997.
[28] L. Ma, H. Huang, P. Zhao, and T. Su,“Acoustic echo cancellation by combining adaptive digitalfilter and recurrent neural network,”CoRR, vol. 2005.09237, 2020.
[29] J.-M. Valin, U. Isik, N. Phansalkar, R. Giri, K. Helwani, and A. Krishnaswamy,“A perceptually-motivated approach for lowcomplexity, real-time enhancement of fullband speech,” CoRR,vol. 2008.04259, 2020.
[30] Y. Hu, Y. Liu, S. Lv, M. Xing, S. Zhang, Y. Fu, J. Wu, B. Zhang,and L. Xie,“Dccrn: Deep complex convolution recurrent network for phase-aware speech enhancement,” CoRR, vol. 2008.00264,2020.
[31] N. L. Westhausen and B. T. Meyer,“Dual-signal transformation lstm network for real-time noise suppression,” CoRR, vol.2005.07551, 2020.
[32] C. K. A. Reddy, V. Gopal, R. Cutler, E. Beyrami, R. Cheng,H. Dubey, S. Matusevych, R. Aichner, A. Aazami, S. Braun,P. Rana, S. Srinivasan, and J. Gehrke,“The interspeech 2020 deep noise suppression challenge: Datasets, subjective testing framework, and challenge results,” 2020.
[33] K. Sridhar, R. Cutler, A. Saabas, T. Parnamaa, M. Loide, H. Gamper, S. Braun, R. Aichner, and S. Srinivasan,“Icassp 2021 acoustic echo cancellation challenge: Datasets, testing framework, and results,” CoRR, vol. 2009.04972, 2020.
[34] C.-C. Lo, S.-W. Fu, W.-C. Huang, X. Wang, J. Yamagishi,Y. Tsao, and H.-M. Wang,“Mosnet: Deep learning based objective assessment for voice conversion,” CoRR, vol. 1904.08352,2019.
[35] I. T. Union,“Itu-t g.168: Digital network echo cancellers.” 2012.[Online]. Available: https://www.itu.int/rec/T-REC-G.168/en
[36] J. L. Ba, J. R. Kiros, and G. E. Hinton,“Layer normalization,”CoRR, vol. 1607.06450, 2016.
[37] D. B. Paul and J. M. Baker,“The design for the Wall Street Journal-based CSR corpus,” in IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), Banff, 1992,pp. 899–902.
[38] I. T. Union,“Subjective performance evaluation of network echo cancellers,” 1998. [Online]. Available: https://www.itu.int/rec/T-REC-P.831/en
[39] M. Frigo and S. G. Johnson,“The design and implementation of FFTW3,” Proceedings of the IEEE, vol. 93, no. 2, pp. 216–231,2005, special issue on“Program Generation, Optimization, and Platform Adaptation”.
[40] J.-M. Valin, S. Tenneti, K. Helwani, U. Isik, and A. Krishnaswamy, “Low-complexity, real-time joint neural echo control and speech enhancement based on percepnet,” CoRR, vol.2102.05245, 2021.

转载请注明:xuhss » 论文翻译:2021_Acoustic Echo Cancellation with Cross-Domain Learning

喜欢 (4)

您必须 登录 才能发表评论!