
{"id":11298,"date":"2026-08-21T18:26:12","date_gmt":"2026-08-21T10:26:12","guid":{"rendered":"https:\/\/infernews.com\/blog\/next-audio-patch-embedding\/"},"modified":"2026-08-21T18:26:12","modified_gmt":"2026-08-21T10:26:12","slug":"next-audio-patch-embedding","status":"publish","type":"post","link":"https:\/\/infernews.com\/blog\/next-audio-patch-embedding\/","title":{"rendered":"NAPE \u7c21\u5316\u81ea\u76e3\u7763\u97f3\u8a0a\u7247\u6bb5\u8a13\u7df4"},"content":{"rendered":"\n<figure class=\"wp-block-image size-large\"><img decoding=\"async\" src=\"https:\/\/infernews.com\/blog\/wp-content\/uploads\/2026\/08\/pasted-b01cfca4195d.jpg\" alt=\"NAPE Architecture\"><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">\u97f3\u8a0a\u6a21\u578b\u8981\u517c\u9867\u8a13\u7df4\u6210\u672c\u3001\u8868\u5fb5\u80fd\u529b\u8207\u6a21\u578b\u898f\u6a21\uff0c\u5f80\u5f80\u9700\u8981\u52a0\u5165\u591a\u500b\u8f14\u52a9\u6a21\u7d44\u3002NAPE\uff08Next Audio Patch Embedding prediction\uff09\u662f\u4e00\u500b\u81ea\u76e3\u7763\u97f3\u8a0a\u8868\u5fb5\u5b78\u7fd2\u6846\u67b6\uff0c\u5c07 log-mel spectrogram \u5207\u6210 patch\uff0c\u518d\u7531\u56e0\u679c Transformer \u6839\u64da\u524d\u9762\u7684\u7247\u6bb5\u9810\u6e2c\u4e0b\u4e00\u500b patch embedding\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">NAPE \u6c92\u6709 reconstruction decoder\u3001acoustic tokenizer\u3001student-teacher \u67b6\u69cb\u6216\u984d\u5916\u6b63\u5247\u5316\u640d\u5931\u3002\u5b83\u4f9d\u9760\u4e09\u500b\u6a5f\u5236\u7dad\u6301\u5b78\u7fd2\u8a0a\u865f\uff1acausal masking \u96b1\u85cf\u672a\u4f86\u4f4d\u7f6e\u3001prediction shift \u8981\u6c42\u4f4d\u7f6e t \u9810\u6e2c t+1\uff0c\u4ee5\u53ca stop-gradient \u56fa\u5b9a\u76ee\u6a19 embedding\uff0c\u907f\u514d\u6a21\u578b\u9000\u5316\u6210\u8f38\u51fa\u76f8\u540c\u5411\u91cf\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e8c\u7dad spectrogram \u6703\u6309\u6307\u5b9a scanning order \u8f49\u6210\u4e00\u7dad\u5e8f\u5217\uff0c\u7814\u7a76\u6db5\u84cb raster\u3001diagonal\u3001zigzag \u548c time-major \u56db\u7a2e\u6392\u5217\uff1b\u5176\u4e2d raster\u3001diagonal \u53ca\u6642\u9593\u65b9\u5411\u7684\u6392\u5217\u8f03\u7b26\u5408\u8072\u97f3\u4e8b\u4ef6\u7684\u767c\u5c55\u3002\u6a21\u578b\u5728 AudioSet \u9810\u8a13\u7df4\uff0c\u518d\u65bc AudioSet-2M\u3001AudioSet-20K\u3001ESC-50\u3001Speech Commands V1\/V2 \u548c IEMOCAP \u9032\u884c\u5fae\u8abf\u6216 linear probing\uff0c\u8cc7\u6599\u986f\u793a\u5b83\u5728\u591a\u9805\u4efb\u52d9\u53d6\u5f97 state-of-the-art fine-tuning \u7d50\u679c\uff0c\u4e26\u5177\u5099\u7a69\u5b9a\u7684\u8de8\u898f\u6a21\u64f4\u5c55\u80fd\u529b\uff0c\u4f46\u539f\u59cb\u8cc7\u8a0a\u6c92\u6709\u63d0\u4f9b\u5404\u9805\u5177\u9ad4\u5206\u6578\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8981\u6c42\u7cfb\u7d71\u6709 Python 3.10\u3001PyTorch 2.8.0 \u548c Transformers 4.56.2 \u7684\u74b0\u5883\uff0c\u4e26\u63d0\u4f9b requirements \u6a94\u53ca\u90e8\u5206\u8cc7\u6599\u96c6\u8655\u7406\u7a0b\u5f0f\uff1bESC-50\u3001Speech Commands V1\/V2 \u548c IEMOCAP \u6709\u4e0b\u8f09\u8173\u672c\uff0cAudioSet \u5247\u8981\u81ea\u884c\u8655\u7406\u6d89\u53ca YouTube \u7684\u4e0b\u8f09\u6d41\u7a0b\u3002\u7814\u7a76\u5718\u968a\u4ea6\u5217\u51fa\u7a0b\u5f0f\u78bc\u53ca\u9810\u8a13\u7df4 checkpoint \u7684\u767c\u5e03\u8cc7\u8a0a\uff0c\u4f46 Hugging Face checkpoint \u4ecd\u6a19\u793a\u70ba\u5f85\u8fa6\uff0c\u4e0d\u80fd\u628a\u5b8c\u6574\u6a21\u578b\u53d6\u5f97\u6d41\u7a0b\u8996\u70ba\u5df2\u7d93\u9f4a\u5099\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u8a13\u7df4\u8a0a\u865f\u7cbe\u7c21<\/strong>\uff1a\u53ea\u7528\u4e0b\u4e00\u500b patch embedding \u9810\u6e2c\u8207 stop-gradient\u3002<\/li>\n<li><strong>\u9069\u5408\u97f3\u8a0a\u8868\u5fb5<\/strong>\uff1a\u53ef\u652f\u63f4\u8a9e\u97f3\u6307\u4ee4\u3001\u74b0\u5883\u8072\u97f3\u53ca\u60c5\u7dd2\u8fa8\u8b58\u7b49\u4efb\u52d9\u3002<\/li>\n<li><strong>\u6392\u5217\u9806\u5e8f\u6709\u5f71\u97ff<\/strong>\uff1aspectrogram \u7684\u7dda\u6027\u5316\u65b9\u5f0f\u6703\u6539\u8b8a\u6a21\u578b\u770b\u5230\u7684\u6642\u9593\u95dc\u4fc2\u3002<\/li>\n<li><strong>\u6e2c\u8a66\u9580\u6abb\u6e05\u695a<\/strong>\uff1a\u9700\u8981 AudioSet \u6216\u4e0b\u6e38\u8cc7\u6599\u96c6\uff0c\u4ee5\u53ca W&amp;B \u8ffd\u8e64\u5be6\u9a57\u3002<\/li>\n<li><strong>\u9650\u5236\u5728\u8cc7\u6599\u6d41\u7a0b<\/strong>\uff1aAudioSet \u4e0d\u63d0\u4f9b\u540c\u7b49\u7c21\u5316\u7684\u4e0b\u8f09\u65b9\u5f0f\uff0ccheckpoint \u53d6\u5f97\u72c0\u614b\u4ea6\u672a\u5b8c\u5168\u660e\u78ba\u3002<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\"><a href=\"https:\/\/umbertocappellazzo.github.io\/nape\/\" rel=\"noopener noreferrer\"><strong>\u9805\u76ee\u4e3b\u9801<\/strong><\/a> \u00b7 <a href=\"https:\/\/github.com\/umbertocappellazzo\/nape\" rel=\"noopener noreferrer\"><strong>GitHub<\/strong><\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>NAPE \u4ee5\u56e0\u679c Transformer \u9810\u6e2c\u4e0b\u4e00\u500b\u97f3\u8a0a patch embedding\uff0c\u7701\u53bb\u89e3\u78bc\u5668\u8207 tokenizer\uff0c\u63a2\u7d22\u66f4\u7cbe\u7c21\u7684\u97f3\u8a0a\u8868\u5fb5\u5b78\u7fd2\u65b9\u6cd5\u3002<\/p>\n","protected":false},"author":8,"featured_media":11297,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"wpai_generated_summary":"","footnotes":""},"categories":[133,29,131,152,127,128],"tags":[],"class_list":["post-11298","post","type-post","status-publish","format-standard","hentry","category-133","category-audio","category-embedding","category-python","category-127","category-128"],"_links":{"self":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/11298","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/users\/8"}],"replies":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/comments?post=11298"}],"version-history":[{"count":0,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/11298\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media\/11297"}],"wp:attachment":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media?parent=11298"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/categories?post=11298"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/tags?post=11298"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}