{"id":4257,"date":"2025-10-16T13:21:15","date_gmt":"2025-10-16T05:21:15","guid":{"rendered":"https:\/\/www.15zhi.net\/blog\/?p=4257"},"modified":"2025-10-20T11:03:29","modified_gmt":"2025-10-20T03:03:29","slug":"202510-%e8%ae%ba%e6%96%87%e7%a0%94%e8%af%bb-lutin-efficient-neural-network-inference-with-table-lookup","status":"publish","type":"post","link":"https:\/\/www.15zhi.net\/blog\/202510-%e8%ae%ba%e6%96%87%e7%a0%94%e8%af%bb-lutin-efficient-neural-network-inference-with-table-lookup\/","title":{"rendered":"202510 \u8bba\u6587\u7814\u8bfb-DiffWave: A Versatile Diffusion Model for Audio Synthesis"},"content":{"rendered":"\n<p>\u4f5c\u8005\uff1aZhifeng Kong\uff0cWei Ping\uff0cJiaji Huang<\/p>\n\n\n\n<p>\u6765\u6e90\uff1aICLR 2021<\/p>\n\n\n\n<p>\u5355\u4f4d\uff1aUniversity of California, San Diego<\/p>\n\n\n\n<p>\u65f6\u95f4\uff1a2021\u5e745\u6708<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u80cc\u666f<\/h2>\n\n\n\n<p>\u5728\u8bed\u97f3\u5408\u6210\u9886\u57df\uff0c\u6df1\u5ea6\u751f\u6210\u6a21\u578b\u5df2\u7ecf\u80fd\u591f\u5408\u6210\u9ad8\u4fdd\u771f\u8bed\u97f3\u3002\u65e9\u671f\u7684\u4e3b\u6d41\u65b9\u6cd5\u5305\u62ec\u81ea\u56de\u5f52\u6a21\u578b\uff08Autoregressive Models\uff09\uff0c\u5982 <em>WaveNet<\/em>\uff0c\u80fd\u9010\u70b9\u751f\u6210\u6ce2\u5f62\uff0c\u4f46\u901f\u5ea6\u6781\u6162\uff1b\u6d41\u6a21\u578b\uff08Flow-based Models\uff09\uff0c\u5982 <em>WaveFlow<\/em>\uff0c\u80fd\u5e76\u884c\u751f\u6210\uff0c\u4f46\u7ed3\u6784\u590d\u6742\u3001\u53c2\u6570\u91cf\u5927\uff1bVAE\u6a21\u578b\uff08Variational Autoencoder\uff09\uff0c\u4f46\u751f\u6210\u8bed\u97f3\u7684\u7ec6\u8282\u8d28\u91cf\u8f83\u4f4e\u3002\u8fd9\u4e9b\u6a21\u578b\u666e\u904d\u5b58\u5728\u8bad\u7ec3\u56f0\u96be\u3001\u63a8\u7406\u6162\u6216\u97f3\u8d28\u5dee\u7684\u95ee\u9898\u3002\u6269\u6563\u6a21\u578b\u6700\u65e9\u5e94\u7528\u4e8e\u56fe\u50cf\u751f\u6210\uff0c\u901a\u8fc7\u201c\u9010\u6b65\u53bb\u566a\u201d\u7684\u65b9\u5f0f\u5c06\u968f\u673a\u566a\u58f0\u8f6c\u5316\u4e3a\u771f\u5b9e\u6837\u672c\u3002\u8fd9\u7bc7\u6587\u7ae0\u9996\u6b21\u5c06\u8fd9\u4e00\u601d\u60f3\u5f15\u5165\u8bed\u97f3\u9886\u57df\uff0c\u63d0\u51fa\u4e86\u4e00\u79cd\u975e\u81ea\u56de\u5f52\u3001\u53ef\u5e76\u884c\u3001\u9ad8\u4fdd\u771f\u7684\u8bed\u97f3\u751f\u6210\u6a21\u578b\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e3b\u8981\u8d21\u732e<\/h2>\n\n\n\n<p>1.\u63d0\u51fa\u9996\u4e2a\u6269\u6563\u6a21\u578b\u7684\u8bed\u97f3\u751f\u6210\u6846\u67b6\uff08DiffWave\uff09\uff1a\u80fd\u5b9e\u73b0\u975e\u81ea\u56de\u5f52\u3001\u5e76\u884c\u7684\u6ce2\u5f62\u5408\u6210\u3002\u517c\u987e\u6709\u6761\u4ef6\u4e0e\u65e0\u6761\u4ef6\u8bed\u97f3\u751f\u6210\uff0c\u5728\u65e0\u6761\u4ef6\u8bed\u97f3\u751f\u6210\u4e0e\u7c7b\u522b\u6761\u4ef6\u751f\u6210\u4efb\u52a1\u4e2d\uff0cDiffWave \u7684\u97f3\u8d28\u4e0e\u6837\u672c\u591a\u6837\u6027\u5747\u663e\u8457\u8d85\u8d8a WaveGAN \u4e0e WaveNet\u3002<\/p>\n\n\n\n<p>2.\u63d0\u51fa\u4e86\u57fa\u4e8e\u524d\u9988\u5f0f\u53cc\u5411\u6269\u5f20\u5377\u79ef\u7684\u97f3\u9891\u751f\u6210\u67b6\u6784\uff0c\u7075\u611f\u6765\u6e90\u4e8e WaveNet\uff0c\u4f46\u975e\u81ea\u56de\u5f52\u3002\u5728\u8bed\u97f3\u8d28\u91cf\u4e0a\u4e0e WaveNet \u6301\u5e73\uff0c\u5374\u80fd\u4ee5\u6781\u9ad8\u901f\u5ea6\u751f\u6210\u957f\u97f3\u9891\uff0c\u53ea\u9700\u6781\u5c11\u7684\u53cd\u5411\u6269\u6563\u6b65\u9aa4\u3002<\/p>\n\n\n\n<p>3.\u8f7b\u91cf\u5316\u4e0e\u9ad8\u4fdd\u771f\u5e76\u5b58\uff0c\u5c3d\u7ba1\u901f\u5ea6\u7565\u6162\u4e8e\u90e8\u5206\u6d41\u6a21\u578b\uff0c\u4f46\u6a21\u578b\u53c2\u6570\u66f4\u5c0f\u3001\u8ba1\u7b97\u6548\u7387\u66f4\u9ad8\uff0c\u672a\u6765\u4ecd\u5177\u4f18\u5316\u6f5c\u529b\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u6280\u672f\u65b9\u6cd5<\/h2>\n\n\n\n<p><strong>1.\u603b\u4f53\u601d\u60f3<\/strong><\/p>\n\n\n\n<p>DiffWave \u7684\u6838\u5fc3\u662f\u4e00\u4e2a\u57fa\u4e8e Markov \u94fe\u7684\u6269\u6563\u2014\u53cd\u5411\u6269\u6563\u8fc7\u7a0b\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"812\" height=\"280\" src=\"https:\/\/www.15zhi.net\/blog\/wp-content\/uploads\/2025\/10\/\u56fe\u7247-6.png\" alt=\"\" class=\"wp-image-4295\" srcset=\"https:\/\/www.15zhi.net\/blog\/wp-content\/uploads\/2025\/10\/\u56fe\u7247-6.png 812w, https:\/\/www.15zhi.net\/blog\/wp-content\/uploads\/2025\/10\/\u56fe\u7247-6-300x103.png 300w, https:\/\/www.15zhi.net\/blog\/wp-content\/uploads\/2025\/10\/\u56fe\u7247-6-768x265.png 768w\" sizes=\"auto, (max-width: 812px) 100vw, 812px\" \/><\/figure>\n\n\n\n<p><strong>\u6269\u6563\u8fc7\u7a0b\uff08Diffusion Process\uff09\uff1a<\/strong><br>\u4ece\u771f\u5b9e\u8bed\u97f3\u6570\u636e x0\u5f00\u59cb\uff0c\u9010\u6b65\u52a0\u5165\u9ad8\u65af\u566a\u58f0\uff0c\u5f97\u5230\u4e00\u7cfb\u5217\u4e2d\u95f4\u72b6\u6001 x1,x2,&#8230;,xT\uff0c\u6700\u7ec8\u53d8\u4e3a\u767d\u566a\u58f0\u3002<\/p>\n\n\n\n<p><strong>\u53cd\u5411\u8fc7\u7a0b\uff08Reverse Process\uff09\uff1a<\/strong><br>\u6a21\u578b\u5b66\u4e60\u5982\u4f55\u9010\u6b65\u53bb\u566a\uff0c\u4ece\u968f\u673a\u566a\u58f0 xT \u9010\u6b65\u8fd8\u539f\u51fa\u7ed3\u6784\u5316\u7684\u97f3\u9891\u6ce2\u5f62 x0\u200b\u3002\u6bcf\u4e00\u6b65\u53cd\u5411\u751f\u6210\u8fc7\u7a0b\u7531\u795e\u7ecf\u7f51\u7edc\u03f5\u03b8\u200b(xt\u200b,t) \u9884\u6d4b\u566a\u58f0\u5206\u91cf\u3002<\/p>\n\n\n\n<p><strong>2.\u7f51\u7edc\u7ed3\u6784<\/strong><\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"699\" height=\"473\" src=\"https:\/\/www.15zhi.net\/blog\/wp-content\/uploads\/2025\/10\/\u56fe\u7247-7.png\" alt=\"\" class=\"wp-image-4296\" srcset=\"https:\/\/www.15zhi.net\/blog\/wp-content\/uploads\/2025\/10\/\u56fe\u7247-7.png 699w, https:\/\/www.15zhi.net\/blog\/wp-content\/uploads\/2025\/10\/\u56fe\u7247-7-300x203.png 300w\" sizes=\"auto, (max-width: 699px) 100vw, 699px\" \/><\/figure>\n\n\n\n<p>DiffWave \u7684\u7f51\u7edc\u7ed3\u6784\u91c7\u7528\u4e00\u79cd\u975e\u81ea\u56de\u5f52\u7684\u524d\u9988\u5f0f\u5377\u79ef\u6846\u67b6\uff0c\u7531\u591a\u5c42\u53cc\u5411\u6269\u5f20\u5377\u79ef\u6784\u6210\u7684\u6b8b\u5dee\u5806\u53e0\u7ec4\u6210\uff0c\u6bcf\u5c42\u5305\u542b\u95e8\u63a7\u6fc0\u6d3b\u3001\u6b8b\u5dee\u8fde\u63a5\u4e0e\u8df3\u8dc3\u8fde\u63a5\uff0c\u4ece\u800c\u5728\u4fdd\u8bc1\u7a33\u5b9a\u8bad\u7ec3\u7684\u540c\u65f6\u878d\u5408\u591a\u5c42\u7279\u5f81\u3002<\/p>\n\n\n\n<p>\u4e0eWaveNet\u4e0d\u540c\uff0cDiffWave\u53d6\u6d88\u4e86\u56e0\u679c\u7ea6\u675f\uff0c\u4f7f\u6a21\u578b\u80fd\u591f\u540c\u65f6\u5229\u7528\u524d\u540e\u65f6\u5e8f\u4fe1\u606f\uff0c\u5728\u5e76\u884c\u6761\u4ef6\u4e0b\u751f\u6210\u6574\u6bb5\u6ce2\u5f62\u3002<\/p>\n\n\n\n<p>\u6a21\u578b\u901a\u8fc7\u6b63\u5f26-\u4f59\u5f26\u65f6\u95f4\u6b65\u5d4c\u5165\uff08timestep embedding\uff09\u611f\u77e5\u4e0d\u540c\u6269\u6563\u9636\u6bb5\u7684\u7279\u5f81\uff0c\u5e76\u53ef\u6839\u636e\u4efb\u52a1\u7c7b\u578b\u5f15\u5165mel\u9891\u8c31\uff08\u5c40\u90e8\u6761\u4ef6\uff09\u6216\u7c7b\u522b\u5d4c\u5165\uff08\u5168\u5c40\u6761\u4ef6\uff09\u4f5c\u4e3a\u751f\u6210\u5f15\u5bfc\u3002\u6574\u4f53\u7ed3\u6784\u8f7b\u91cf\u9ad8\u6548\uff0c\u4ec5\u9700\u5c11\u91cf\u53c2\u6570\u5373\u53ef\u751f\u621022.05 kHz\u9ad8\u4fdd\u771f\u8bed\u97f3\uff0c\u517c\u987e\u97f3\u8d28\u4e0e\u63a8\u7406\u901f\u5ea6\uff0c\u5c55\u73b0\u4e86\u6269\u6563\u6a21\u578b\u5728\u97f3\u9891\u5efa\u6a21\u4e2d\u7684\u4f18\u8d8a\u6027\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u5b9e\u9a8c\u5206\u6790<\/h2>\n\n\n\n<p>\u4f5c\u8005\u8bbe\u8ba1\u4e86\u4e09\u7c7b\u5b9e\u9a8c\u4efb\u52a1\uff0c\u4ee5\u9a8c\u8bc1 DiffWave \u6a21\u578b\u5728\u4e0d\u540c\u97f3\u9891\u751f\u6210\u573a\u666f\u4e0b\u7684\u6027\u80fd\u4e0e\u901a\u7528\u6027\u3002<\/p>\n\n\n\n<p><strong>1.\u6761\u4ef6\u8bed\u97f3\u751f\u6210<\/strong><\/p>\n\n\n\n<p>\u6570\u636e\u96c6\uff1aLJSpeech \u6570\u636e\u96c6\u5305\u542b\u5355\u4e00\u5973\u58f0\u6717\u8bfb 13,100 \u6bb5\u77ed\u8bed\u97f3\uff0c\u65f6\u957f\u7ea6 24 \u5c0f\u65f6\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"790\" height=\"306\" src=\"https:\/\/www.15zhi.net\/blog\/wp-content\/uploads\/2025\/10\/\u56fe\u7247-8.png\" alt=\"\" class=\"wp-image-4299\" srcset=\"https:\/\/www.15zhi.net\/blog\/wp-content\/uploads\/2025\/10\/\u56fe\u7247-8.png 790w, https:\/\/www.15zhi.net\/blog\/wp-content\/uploads\/2025\/10\/\u56fe\u7247-8-300x116.png 300w, https:\/\/www.15zhi.net\/blog\/wp-content\/uploads\/2025\/10\/\u56fe\u7247-8-768x297.png 768w\" sizes=\"auto, (max-width: 790px) 100vw, 790px\" \/><\/figure>\n\n\n\n<p>\u5b9e\u9a8c\u7ed3\u679c\u8868\u660e\uff0cDiffWave \u5728\u97f3\u8d28\u4e0a\u4e0e WaveNet \u51e0\u4e4e\u6301\u5e73\uff0c\u4f46\u63a8\u7406\u901f\u5ea6\u63d0\u5347\u4e86\u4e0a\u767e\u500d\u3002DiffWave\u7684\u53c2\u6570\u91cf\u66f4\u5c0f\uff0c\u4ec5\u4e3a WaveGlow \u7684 1\/30\u3002\u5728\u672a\u8fdb\u884c CUDA \u5185\u6838\u4f18\u5316\u7684\u60c5\u51b5\u4e0b\u5373\u53ef\u8d85\u5b9e\u65f6\u751f\u6210\uff0c\u663e\u793a\u51fa\u6781\u5f3a\u7684\u8ba1\u7b97\u6548\u7387\u3002<\/p>\n\n\n\n<p><strong>2.\u65e0\u6761\u4ef6\u8bed\u97f3\u751f\u6210<\/strong><\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"731\" height=\"175\" src=\"https:\/\/www.15zhi.net\/blog\/wp-content\/uploads\/2025\/10\/\u56fe\u7247-9.png\" alt=\"\" class=\"wp-image-4300\" srcset=\"https:\/\/www.15zhi.net\/blog\/wp-content\/uploads\/2025\/10\/\u56fe\u7247-9.png 731w, https:\/\/www.15zhi.net\/blog\/wp-content\/uploads\/2025\/10\/\u56fe\u7247-9-300x72.png 300w\" sizes=\"auto, (max-width: 731px) 100vw, 731px\" \/><\/figure>\n\n\n\n<p>DiffWave \u5728\u5b8c\u5168\u65e0\u6761\u4ef6\u60c5\u51b5\u4e0b\u4ecd\u80fd\u751f\u6210\u81ea\u7136\u5ea6\u8f83\u9ad8\u7684\u8bed\u97f3\u6ce2\u5f62\uff0c\u663e\u8457\u8d85\u8d8a WaveGAN \u4e0e WaveNet\uff0c\u8bc1\u660e\u5176\u5f3a\u5927\u7684\u6982\u7387\u5efa\u6a21\u80fd\u529b\u3002<\/p>\n\n\n\n<p><strong>3.\u7c7b\u522b\u6761\u4ef6\u8bed\u97f3\u751f\u6210<\/strong><\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"776\" height=\"179\" src=\"https:\/\/www.15zhi.net\/blog\/wp-content\/uploads\/2025\/10\/\u56fe\u7247-10.png\" alt=\"\" class=\"wp-image-4301\" srcset=\"https:\/\/www.15zhi.net\/blog\/wp-content\/uploads\/2025\/10\/\u56fe\u7247-10.png 776w, https:\/\/www.15zhi.net\/blog\/wp-content\/uploads\/2025\/10\/\u56fe\u7247-10-300x69.png 300w, https:\/\/www.15zhi.net\/blog\/wp-content\/uploads\/2025\/10\/\u56fe\u7247-10-768x177.png 768w\" sizes=\"auto, (max-width: 776px) 100vw, 776px\" \/><\/figure>\n\n\n\n<p>DiffWave \u5728\u7c7b\u522b\u6761\u4ef6\u4efb\u52a1\u4e2d\u751f\u6210\u7684\u8bed\u97f3\u4e0d\u4ec5\u97f3\u8d28\u81ea\u7136\uff0c\u800c\u4e14\u7c7b\u522b\u8fa8\u8bc6\u5ea6\u66f4\u9ad8\uff0c\u8bf4\u660e\u5176\u5728\u6709\u6761\u4ef6\u751f\u6210\u4e2d\u7684\u4e00\u81f4\u6027\u4e0e\u53ef\u63a7\u6027\u4f18\u5f02\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u603b\u7ed3<\/h2>\n\n\n\n<p>DiffWave \u5c55\u793a\u4e86\u6269\u6563\u6a21\u578b\u5728\u8bed\u97f3\u751f\u6210\u9886\u57df\u7684\u5f3a\u5927\u6f5c\u529b\uff0c\u5145\u5206\u8bc1\u660e\u4e86\u5176\u65b9\u6cd5\u5728\u751f\u6210\u8d28\u91cf\u4e0e\u8ba1\u7b97\u6548\u7387\u4e0a\u7684\u7a81\u7834\u6027\u4f18\u52bf\u3002\u5b83\u4e0d\u4ec5\u80fd\u5b9e\u73b0\u9ad8\u4fdd\u771f\u3001\u9ad8\u901f\u5ea6\u7684\u8bed\u97f3\u5408\u6210\uff0c\u5e76\u4e14\u5728\u65e0\u6761\u4ef6\u548c\u7c7b\u522b\u6761\u4ef6\u751f\u6210\u4efb\u52a1\u4e2d\uff0c\u6a21\u578b\u751f\u6210\u7684\u8bed\u97f3\u4e0d\u4ec5\u81ea\u7136\u6d41\u7545\uff0c\u800c\u4e14\u6837\u672c\u591a\u6837\u6027\u66f4\u9ad8\u3002\u603b\u4f53\u800c\u8a00\uff0cDiffWave \u4e0d\u4ec5\u8bc1\u660e\u4e86\u6269\u6563\u6a21\u578b\u5728\u8fde\u7eed\u65f6\u95f4\u4fe1\u53f7\u5efa\u6a21\u4e2d\u7684\u53ef\u884c\u6027\u4e0e\u4f18\u8d8a\u6027\uff0c\u4e5f\u4e3a\u540e\u7eed\u8bed\u97f3\u6269\u6563\u6a21\u578b\u5960\u5b9a\u4e86\u7406\u8bba\u4e0e\u65b9\u6cd5\u57fa\u7840\uff0c\u5f00\u542f\u4e86\u9ad8\u4fdd\u771f\u3001\u9ad8\u6548\u7387\u3001\u53ef\u63a7\u8bed\u97f3\u751f\u6210\u7684\u65b0\u65b9\u5411\u3002<\/p>\n\n\n\n<p><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u4f5c\u8005\uff1aZhifeng Kong\uff0cWei Ping\uff0cJiaji Huang \u6765\u6e90\uff1aICLR 2021 \u5355\u4f4d\uff1aUn [&hellip;]<\/p>\n","protected":false},"author":72,"featured_media":0,"comment_status":"closed","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[],"class_list":["post-4257","post","type-post","status-publish","format-standard","hentry","category-events"],"_links":{"self":[{"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/posts\/4257","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/users\/72"}],"replies":[{"embeddable":true,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/comments?post=4257"}],"version-history":[{"count":5,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/posts\/4257\/revisions"}],"predecessor-version":[{"id":4303,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/posts\/4257\/revisions\/4303"}],"wp:attachment":[{"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/media?parent=4257"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/categories?post=4257"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/tags?post=4257"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}