{"id":774,"date":"2026-02-24T05:50:37","date_gmt":"2026-02-23T21:50:37","guid":{"rendered":"https:\/\/tastestars.com\/?p=774"},"modified":"2026-02-24T05:50:37","modified_gmt":"2026-02-23T21:50:37","slug":"%e6%8e%a2%e7%b4%a2%e4%b8%8d%e5%90%8cdecoder-head%e5%9c%a8midi-%e5%a4%9a%e5%b1%9e%e6%80%a7%e5%ba%8f%e5%88%97%e7%bb%93%e7%a0%81%e7%9a%84%e5%b7%a5%e4%bd%9c","status":"publish","type":"post","link":"https:\/\/tastestars.com\/index.php\/2026\/02\/24\/%e6%8e%a2%e7%b4%a2%e4%b8%8d%e5%90%8cdecoder-head%e5%9c%a8midi-%e5%a4%9a%e5%b1%9e%e6%80%a7%e5%ba%8f%e5%88%97%e7%bb%93%e7%a0%81%e7%9a%84%e5%b7%a5%e4%bd%9c\/","title":{"rendered":"\u63a2\u7d22\u4e0d\u540cdecoder head\u5728midi \u591a\u5c5e\u6027\u5e8f\u5217\u7ed3\u7801\u7684\u5de5\u4f5c"},"content":{"rendered":"<blockquote><p>\n  \u4e00\u7bc7\u5173\u4e8emidi decoder\u89e3\u7801\u5668\u7684\u79d1\u666e\u5de5\u4f5c\uff0chttps:\/\/github.com\/JudeJiwoo\/nmt\n<\/p><\/blockquote>\n<h3>Nested Music Transformer\uff1a\u5728\u7b26\u53f7\u97f3\u4e50\u4e0e\u97f3\u9891\u751f\u6210\u4e2d\u987a\u5e8f\u89e3\u7801\u590d\u5408 Token<\/h3>\n<p>Jiwoo Ryu\uff0cHao-Wen Dong\uff0cJongmin Jung\uff0cDasaem Jeong<br \/>\n1 \u97e9\u56fd\u897f\u6c5f\u5927\u5b66 \u4eba\u5de5\u667a\u80fd\u7cfb<br \/>\n2 \u52a0\u5dde\u5927\u5b66\u5723\u5730\u4e9a\u54e5\u5206\u6821<br \/>\n3 \u97e9\u56fd\u897f\u6c5f\u5927\u5b66 \u827a\u672f\u4e0e\u6280\u672f\u7cfb<\/p>\n<p>\u6458\u8981<\/p>\n<p>\u4f7f\u7528\u590d\u5408 token \u8868\u793a\u7b26\u53f7\u97f3\u4e50\uff0c\u5373\u6bcf\u4e2a token \u7531\u591a\u4e2a\u4e0d\u540c\u7684\u5b50 token \u7ec4\u6210\uff0c\u6bcf\u4e2a\u5b50 token \u8868\u793a\u4e00\u79cd\u72ec\u7acb\u7684\u97f3\u4e50\u7279\u5f81\u6216\u5c5e\u6027\uff0c\u8fd9\u79cd\u8868\u793a\u65b9\u5f0f\u7684\u4f18\u52bf\u5728\u4e8e\u53ef\u4ee5\u51cf\u5c11\u5e8f\u5217\u957f\u5ea6\u3002\u867d\u7136\u5df2\u6709\u7814\u7a76\u9a8c\u8bc1\u4e86\u590d\u5408 token \u5728\u97f3\u4e50\u5e8f\u5217\u5efa\u6a21\u4e2d\u7684\u6709\u6548\u6027\uff0c\u4f46\u82e5\u540c\u65f6\u9884\u6d4b\u6240\u6709\u5b50 token\uff0c\u53ef\u80fd\u65e0\u6cd5\u5145\u5206\u6355\u6349\u5b83\u4eec\u4e4b\u95f4\u7684\u76f8\u4e92\u4f9d\u8d56\u5173\u7cfb\uff0c\u4ece\u800c\u5bfc\u81f4\u6b21\u4f18\u7ed3\u679c\u3002<\/p>\n<p>\u672c\u6587\u63d0\u51fa\u4e86 Nested Music Transformer\uff08NMT\uff09\uff0c\u8fd9\u662f\u4e00\u79cd\u4e13\u4e3a\u81ea\u56de\u5f52\u89e3\u7801\u590d\u5408 token \u800c\u8bbe\u8ba1\u7684\u67b6\u6784\u3002\u8be5\u67b6\u6784\u7c7b\u4f3c\u4e8e\u5bf9\u5c55\u5e73 token \u7684\u5904\u7406\u65b9\u5f0f\uff0c\u4f46\u5185\u5b58\u6d88\u8017\u8f83\u4f4e\u3002NMT \u7531\u4e24\u4e2a Transformer \u7ec4\u6210\uff1a\u4e00\u4e2a\u4e3b\u89e3\u7801\u5668\u7528\u4e8e\u5efa\u6a21\u590d\u5408 token \u7684\u5e8f\u5217\uff0c\u53e6\u4e00\u4e2a\u5b50\u89e3\u7801\u5668\u7528\u4e8e\u5efa\u6a21\u6bcf\u4e2a\u590d\u5408 token \u5185\u90e8\u7684\u5b50 token\u3002\u5b9e\u9a8c\u7ed3\u679c\u8868\u660e\uff0c\u5728\u591a\u4e2a\u7b26\u53f7\u97f3\u4e50\u6570\u636e\u96c6\u4ee5\u53ca MAESTRO \u6570\u636e\u96c6\u7684\u79bb\u6563\u97f3\u9891 token \u4e0a\uff0c\u91c7\u7528 NMT \u5904\u7406\u590d\u5408 token \u53ef\u4ee5\u5728\u56f0\u60d1\u5ea6\uff08perplexity\uff09\u65b9\u9762\u53d6\u5f97\u66f4\u4f18\u8868\u73b0\u3002<\/p>\n<p><a class=\"wp-editor-md-post-content-link\" href=\"https:\/\/tastestars.com\/wp-content\/uploads\/2026\/02\/wp_editor_md_f5d37b3cd14ec50bb1b1ed737e587def.jpg\"><img decoding=\"async\" src=\"https:\/\/tastestars.com\/wp-content\/uploads\/2026\/02\/wp_editor_md_f5d37b3cd14ec50bb1b1ed737e587def.jpg\" alt=\"\" \/><\/a><\/p>\n<ol>\n<li>\u5f15\u8a00<\/li>\n<\/ol>\n<p>\u81ea\u56de\u5f52\u8bed\u8a00\u6a21\u578b\u5728\u5404\u7c7b\u751f\u6210\u4efb\u52a1\u4e2d\u5df2\u5360\u636e\u4e3b\u5bfc\u5730\u4f4d\uff0c\u5305\u62ec\u97f3\u4e50\u751f\u6210\u3002\u8bed\u8a00\u6a21\u578b\u4e00\u76f4\u662f\u7b26\u53f7\u97f3\u4e50\u751f\u6210\u4e2d\u6700\u5e7f\u6cdb\u4f7f\u7528\u7684\u751f\u6210\u6a21\u578b\u3002\u968f\u7740\u5411\u91cf\u91cf\u5316\u6216\u6b8b\u5dee\u5411\u91cf\u91cf\u5316\u6280\u672f\u7684\u6210\u529f\uff0c\u8bed\u8a00\u6a21\u578b\u4e5f\u88ab\u5e7f\u6cdb\u5e94\u7528\u4e8e\u97f3\u9891\u57df\u97f3\u4e50\u751f\u6210\u3002<\/p>\n<p>\u8bed\u8a00\u6a21\u578b\u7684\u4f18\u52bf\u6765\u6e90\u4e8e\u5176\u5bf9\u5e8f\u5217\u4fe1\u606f\u7684\u81ea\u56de\u5f52\u5efa\u6a21\u80fd\u529b\u3002\u4e00\u65e6\u6570\u636e\u88ab\u5c55\u5e73\u4e3a\u79bb\u6563 token \u5e8f\u5217\uff0c\u8bed\u8a00\u6a21\u578b\u4fbf\u53ef\u76f4\u63a5\u5e94\u7528\u3002\u5df2\u6709\u8bb8\u591a\u5de5\u4f5c\u5c06\u7b26\u53f7\u97f3\u4e50\u8868\u793a\u4e3a\u5c55\u5e73 token \u5e8f\u5217\uff0c\u4f8b\u5982 MIDI-like \u7f16\u7801\u6216 REMI \u7f16\u7801\u3002<\/p>\n<p>\u7136\u800c\uff0c\u8fd9\u79cd\u65b9\u6cd5\u7684\u4e00\u4e2a\u5c40\u9650\u5728\u4e8e\u5e8f\u5217\u957f\u5ea6\u8f83\u957f\u3002\u4f8b\u5982\uff0c\u5728 Lakh MIDI \u6570\u636e\u96c6\u4e2d\uff0c\u4f5c\u54c1\u7684\u5e73\u5747 token \u6570\u91cf\u8fbe\u5230 14,647 \u4e2a\u3002\u4e3a\u4e86\u89e3\u51b3\u8fd9\u4e00\u95ee\u9898\uff0cCompound Word Transformer \u63d0\u51fa\u4e86\u4e00\u79cd\u540d\u4e3a Compound Word \u7684\u7f16\u7801\u65b9\u5f0f\uff0c\u5c06\u591a\u4e2a\u97f3\u4e50\u7279\u5f81\u7f16\u7801\u4e3a\u4e00\u4e2a\u591a\u7ef4\u590d\u5408 token\u3002\u901a\u8fc7\u5c06\u97f3\u4e50\u7279\u5f81\u5206\u4e3a\u201c\u8282\u62cd\u7c7b\u201d\u548c\u201c\u97f3\u7b26\u7c7b\u201d\u4e24\u79cd\u590d\u5408 token \u7c7b\u578b\uff0c\u8be5\u65b9\u6cd5\u5c06\u5e8f\u5217\u957f\u5ea6\u7f29\u77ed\u81f3 REMI \u7684\u4e00\u534a\u4ee5\u4e0b\u3002\u7c7b\u4f3c\u5730\uff0cMultitrack Music Transformer \u5c06\u62cd\u4f4d\u3001\u4e50\u5668\u3001\u97f3\u9ad8\u548c\u65f6\u503c\u7f16\u7801\u4e3a\u5355\u4e00 token\uff0c\u4f7f\u5e8f\u5217\u957f\u5ea6\u7ea6\u4e3a REMI \u7684\u4e09\u5206\u4e4b\u4e00\u3002\u6b64\u5916\uff0c\u97f3\u7b26\u7ea7\u522b\u7684\u590d\u5408 token \u5728\u98ce\u683c\u8bc6\u522b\u6216\u4f34\u594f\u5efa\u8bae\u7b49\u5224\u522b\u4efb\u52a1\u4e2d\u8868\u73b0\u51fa\u660e\u663e\u4f18\u52bf\u3002<\/p>\n<p>\u5c3d\u7ba1\u8fd9\u4e9b\u65b9\u6cd5\u901a\u8fc7\u6253\u5305\u97f3\u4e50\u7279\u5f81\u4ee5\u51cf\u5c11\u5e8f\u5217\u957f\u5ea6\uff0c\u4f46\u5728\u7b26\u53f7\u97f3\u4e50\u751f\u6210\u4efb\u52a1\u4e2d\uff0c\u50cf REMI \u8fd9\u6837\u7684\u5c55\u5e73\u7f16\u7801\u4ecd\u7136\u5360\u4e3b\u5bfc\u5730\u4f4d\u3002\u5728\u90e8\u5206\u7814\u7a76\u7684\u4e3b\u89c2\u542c\u611f\u6d4b\u8bd5\u4e2d\uff0cREMI \u7684\u751f\u6210\u6548\u679c\u66f4\u53d7\u504f\u597d\u3002\u539f\u56e0\u4e4b\u4e00\u5728\u4e8e\uff0c\u73b0\u6709\u6a21\u578b\u5728\u9884\u6d4b\u590d\u5408 token \u65f6\u91c7\u7528\u5e76\u884c\u6216\u90e8\u5206\u987a\u5e8f\u9884\u6d4b\u65b9\u5f0f\uff0c\u800c\u672a\u5145\u5206\u8003\u8651\u590d\u5408 token \u5185\u90e8\u4e0d\u540c\u97f3\u4e50\u7279\u5f81\u4e4b\u95f4\u7684\u4f9d\u8d56\u5173\u7cfb\u3002<\/p>\n<p>\u4e3a\u4e86\u89e3\u51b3\u8fd9\u4e00\u95ee\u9898\uff0c\u6211\u4eec\u63d0\u51fa\u4e86\u4e00\u79cd\u65b0\u7684\u89e3\u7801\u6846\u67b6\u2014\u2014Nested Music Transformer\uff08NMT\uff09\u3002\u8be5\u6846\u67b6\u7684\u4e3b\u8981\u76ee\u6807\u662f\u5728\u4fdd\u6301\u9ad8\u6548\u5185\u5b58\u4f7f\u7528\u7684\u540c\u65f6\uff0c\u4ee5\u5b8c\u5168\u987a\u5e8f\u7684\u65b9\u5f0f\u89e3\u7801\u590d\u5408 token\u3002NMT \u5728\u5176\u5b50\u89e3\u7801\u5668\u4e2d\u7ed3\u5408\u4e86\u4e24\u79cd\u4ea4\u53c9\u6ce8\u610f\u529b\u7ed3\u6784\uff1aintra-token decoder \u548c Embedding Enricher\u3002\u524d\u8005\u7528\u4e8e\u81ea\u56de\u5f52\u5730\u89e3\u7801\u5355\u4e2a\u590d\u5408 token \u5185\u90e8\u7684\u5b50 token\uff0c\u540e\u8005\u901a\u8fc7\u5173\u6ce8\u524d\u5e8f\u590d\u5408 token \u7684\u9690\u85cf\u72b6\u6001\u6765\u66f4\u65b0\u6bcf\u4e2a\u5b50 token \u7684 embedding\u3002<\/p>\n<p>\u6211\u4eec\u8bc1\u660e\uff0c\u6240\u63d0\u51fa\u7684\u67b6\u6784\u5728\u6027\u80fd\u4e0a\u53ef\u4e0e\u5c55\u5e73\u7f16\u7801\u6a21\u578b\u76f8\u5ab2\u7f8e\uff0c\u540c\u65f6\u5728 GPU \u5185\u5b58\u4f7f\u7528\u548c\u8bad\u7ec3\u65f6\u95f4\u65b9\u9762\u66f4\u52a0\u9ad8\u6548\u3002\u8fd9\u4e00\u7ed3\u8bba\u901a\u8fc7\u7b26\u53f7\u97f3\u4e50\u751f\u6210\u7684\u5b9a\u91cf\u8bc4\u4f30\u4e0e\u4e3b\u89c2\u542c\u611f\u6d4b\u8bd5\u5f97\u5230\u9a8c\u8bc1\u3002\u6b64\u5916\uff0c\u6211\u4eec\u8fd8\u5728\u79bb\u6563\u97f3\u9891 token \u4e0a\u8fdb\u884c\u4e86\u5b9e\u9a8c\uff0c\u7ed3\u679c\u8868\u660e NMT \u53ca\u5176\u4ed6\u5d4c\u5957\u7ed3\u6784\u5728\u97f3\u9891\u751f\u6210\u4efb\u52a1\u4e0a\u4e5f\u80fd\u8fbe\u5230\u4e0e\u5f3a\u57fa\u7ebf\u6a21\u578b\u76f8\u5f53\u7684\u8868\u73b0\u3002<\/p>\n<ol start=\"2\">\n<li>\u57fa\u4e8e\u97f3\u7b26\u7684\u7f16\u7801\uff08Note-Based Encoding\uff09<\/li>\n<\/ol>\n<p>\u5728\u4ecb\u7ecd Nested Music Transformer \u4e4b\u524d\uff0c\u6211\u4eec\u9996\u5148\u8bf4\u660e\u672c\u6587\u91c7\u7528\u7684\u590d\u5408 token \u7f16\u7801\u65b9\u5f0f\u2014\u2014Note-based encoding\uff08NB\uff09\u3002NB \u7684\u7279\u70b9\u662f\u80fd\u591f\u5728\u5355\u4e2a\u590d\u5408 token \u4e2d\u5305\u542b\u6700\u5168\u9762\u7684\u97f3\u4e50\u7279\u5f81\u3002<\/p>\n<p>REMI\u3001Compound Word \u548c NB \u90fd\u4f7f\u7528\u591a\u79cd\u97f3\u4e50\u7279\u5f81\u6765\u8868\u793a\u97f3\u4e50\u3002\u672c\u6587\u5171\u4f7f\u7528\u516b\u79cd\u7279\u5f81\uff1a\u62cd\u4f4d\u3001\u97f3\u9ad8\u548c\u65f6\u503c\u4e3a\u57fa\u672c\u7279\u5f81\uff0c\u4e50\u5668\u3001\u548c\u5f26\u3001\u901f\u5ea6\u548c\u529b\u5ea6\u6839\u636e\u6570\u636e\u96c6\u7279\u6027\u9009\u62e9\u6027\u4f7f\u7528\u3002\u6b64\u5916\uff0c\u8fd8\u4f7f\u7528\u4e00\u4e2a\u989d\u5916\u7279\u5f81 Metric\uff08\u6216 Type\uff09\u6765\u8868\u793a\u5c0f\u8282\u8fb9\u754c\u6216\u62cd\u53f7\u53d8\u5316\u3002<\/p>\n<p>\u5728 Compound Word \u4e2d\uff0c\u97f3\u4e50\u7279\u5f81\u88ab\u5206\u4e3a\u201c\u8282\u62cd\u7c7b\u201d\u548c\u201c\u97f3\u7b26\u7c7b\u201d\u4e24\u7ec4\uff0c\u5e76\u901a\u8fc7 Type token \u6307\u793a\u7c7b\u522b\u3002\u800c NB \u4e0d\u9700\u8981\u8fd9\u79cd\u5206\u7ec4\u6807\u8bc6\u3002\u7531\u4e8e NB \u4e2d\u6bcf\u4e2a\u97f3\u7b26 token \u90fd\u5e26\u6709\u62cd\u4f4d\u4fe1\u606f\uff0c\u56e0\u6b64\u6211\u4eec\u8bbe\u8ba1\u4e86 Metric \u7279\u5f81\u6765\u7f16\u7801\u8282\u62cd\u7ed3\u6784\u53d8\u5316\u3002Metric \u6307\u793a\u5f53\u524d\u97f3\u7b26\u662f\u5426\u5f15\u5165\u65b0\u7684\u62cd\u53f7\u3001\u5c0f\u8282\u6216\u62cd\u4f4d\uff0c\u6216\u5ef6\u7eed\u524d\u4e00\u8282\u62cd\u7ed3\u6784\u3002\u6211\u4eec\u5b9a\u4e49\u4e86\u56db\u79cd\u4e0d\u540c\u7684 Metric \u53d6\u503c\u4ee5\u8868\u793a\u4e0d\u540c\u7684\u8282\u62cd\u53d8\u5316\u7ec4\u5408\u3002<\/p>\n<p>Beat \u8868\u793a\u97f3\u7b26\u5728\u5c0f\u8282\u5185\u7684\u76f8\u5bf9\u4f4d\u7f6e\u3002Chord \u901a\u8fc7\u89c4\u5219\u7b97\u6cd5\u83b7\u5f97\u3002Tempo \u91c7\u7528\u6307\u6570\u523b\u5ea6\u8868\u793a\u901f\u5ea6\u53d8\u5316\u3002Instrument \u6307\u793a\u6f14\u594f\u97f3\u7b26\u7684\u4e50\u5668\uff0c\u9650\u5236\u4e3a 61 \u7c7b\u3002Pitch \u4f7f\u7528 MIDI \u7684 128 \u79cd\u97f3\u9ad8\u3002Duration \u8868\u793a\u97f3\u7b26\u6301\u7eed\u65f6\u95f4\u3002Velocity \u8868\u793a MIDI \u529b\u5ea6\u3002<\/p>\n<p>\u5bf9\u4e8e NB \u7f16\u7801\uff0c\u4e00\u9996\u5305\u542b K \u4e2a\u97f3\u7b26\u7684\u4e50\u66f2\u53ef\u4ee5\u8868\u793a\u4e3a K \u4e2a\u590d\u5408 token \u5e8f\u5217\uff0c\u6bcf\u4e2a token \u5305\u542b\u5982\u4e0b\u5b50 token\uff1a<\/p>\n<p>(metric, beat, chord, tempo, instrument, pitch, duration, velocity)<\/p>\n<p>2.2 \u590d\u5408\u79fb\u4f4d\uff08Compound Shift\uff09<\/p>\n<p>\u901a\u8fc7\u91cd\u65b0\u6392\u5217\u5b50 token \u7684\u987a\u5e8f\uff0c\u53ef\u4ee5\u5c06\u76ee\u6807\u5b50 token \u653e\u5728\u9996\u4f4d\u9884\u6d4b\uff0c\u4ece\u800c\u4f7f\u5176\u66f4\u591a\u5730\u4f9d\u8d56\u4e3b\u89e3\u7801\u5668\u800c\u975e\u5b50\u89e3\u7801\u5668\u3002\u4f8b\u5982\uff0c\u5728 pitch-first \u65b9\u6848\u4e2d\uff0c\u4e00\u4e2a\u590d\u5408 token \u7684\u6392\u5217\u4e3a\uff1a<\/p>\n<p>(pitch, duration, velocity, metric, beat, chord, tempo, instrument)<\/p>\n<p>\u9700\u8981\u6ce8\u610f\u7684\u662f\uff0c\u6574\u4f53\u5e8f\u5217\u7684\u9884\u6d4b\u987a\u5e8f\u4e0d\u53d8\uff0c\u4ec5\u6539\u53d8\u590d\u5408 token \u5185\u90e8\u7684\u5206\u7ec4\u8fb9\u754c\u3002\u672a\u79fb\u4f4d\u8868\u793a\u4e3a NB-MF\uff08metric-first\uff09\uff0c\u79fb\u4f4d\u7248\u672c\u79f0\u4e3a NB-PF\uff08pitch-first\uff09\u3002<\/p>\n<ol start=\"3\">\n<li>Nested Music Transformer<\/li>\n<\/ol>\n<p><a class=\"wp-editor-md-post-content-link\" href=\"https:\/\/tastestars.com\/wp-content\/uploads\/2026\/02\/wp_editor_md_a802dd2987d84d5413c1b5017dd5975a.jpg\"><img decoding=\"async\" src=\"https:\/\/tastestars.com\/wp-content\/uploads\/2026\/02\/wp_editor_md_a802dd2987d84d5413c1b5017dd5975a.jpg\" alt=\"\" \/><\/a><br \/>\nNMT \u5305\u542b\u4e09\u4e2a\u4e3b\u8981\u7ec4\u4ef6\uff1atoken embedding\u3001\u4e3b\u89e3\u7801\u5668\u548c\u5b50\u89e3\u7801\u5668\u3002<\/p>\n<p>\u9996\u5148\uff0c\u5c06\u6bcf\u4e2a\u5b50 token \u7684 embedding \u5728\u5b50 token \u7ef4\u5ea6\u4e0a\u6c42\u548c\uff0c\u5e76\u52a0\u5165\u53ef\u5b66\u4e60\u7684\u7edd\u5bf9\u4f4d\u7f6e\u7f16\u7801\uff0c\u5f97\u5230\u590d\u5408 token \u7684\u8868\u793a\u3002\u7136\u540e\u5c06\u5176\u8f93\u5165\u4e3b\u89e3\u7801\u5668\uff08decoder-only Transformer\uff09\uff0c\u5f97\u5230\u9690\u85cf\u5411\u91cf hi\u3002<\/p>\n<p>\u5b50\u89e3\u7801\u5668\u7684\u76ee\u6807\u662f\u5728\u7ed9\u5b9a\u4e3b\u89e3\u7801\u5668\u8f93\u51fa hi \u548c\u5df2\u751f\u6210\u5b50 token \u7684\u60c5\u51b5\u4e0b\uff0c\u9884\u6d4b\u5f53\u524d\u590d\u5408 token \u7684\u4e0b\u4e00\u4e2a\u5b50 token\u3002<\/p>\n<p>\u4e0e\u4ee5\u5f80\u4f7f\u7528\u62fc\u63a5\u3001RNN \u6216\u81ea\u6ce8\u610f\u529b\u7684\u65b9\u6cd5\u4e0d\u540c\uff0c\u6211\u4eec\u53d1\u73b0\u4f7f\u7528\u4ea4\u53c9\u6ce8\u610f\u529b\u662f\u5efa\u6a21\u590d\u5408 token \u5e8f\u5217\u7684\u6709\u6548\u65b9\u5f0f\u3002\u5177\u4f53\u800c\u8a00\uff0c\u5b50\u89e3\u7801\u5668\u901a\u8fc7\u8fed\u4ee3\u65b9\u5f0f\u6784\u5efa key\/value \u5e8f\u5217\uff0c\u521d\u59cb\u4e3a BOS token\u3002\u5bf9\u4e8e\u7b2c j \u4e2a\u5b50 token\uff0c\u5176\u67e5\u8be2\u5411\u91cf\u4e3a\u7ecf\u8fc7\u4f4d\u7f6e\u7f16\u7801\u7684 hi\uff0ckey\/value \u4e3a BOS \u52a0\u4e0a\u5df2\u751f\u6210\u5b50 token \u7684 embedding\u3002\u901a\u8fc7\u4ea4\u53c9\u6ce8\u610f\u529b\u8ba1\u7b97\u540e\uff0c\u518d\u901a\u8fc7\u7ebf\u6027\u53d8\u6362\u4e0e softmax \u5f97\u5230\u9884\u6d4b\u6982\u7387\u3002<\/p>\n<p>Embedding Enricher<\/p>\n<p>\u7531\u4e8e\u5b50 token embedding \u76f8\u5bf9\u6d45\u5c42\uff0c\u6211\u4eec\u8bbe\u8ba1\u4e86 Embedding Enricher\uff0c\u901a\u8fc7\u5bf9\u8fc7\u53bb w \u4e2a\u4e3b\u89e3\u7801\u5668\u9690\u85cf\u72b6\u6001\u8fdb\u884c\u4ea4\u53c9\u6ce8\u610f\u529b\uff0c\u589e\u5f3a\u5b50 token embedding \u7684\u8bed\u4e49\u8868\u8fbe\u3002\u5b9e\u9a8c\u8868\u660e\uff0c\u8fd9\u79cd\u589e\u5f3a\u673a\u5236\u5728\u7b26\u53f7\u97f3\u4e50\u4efb\u52a1\u4e2d\u5e26\u6765\u66f4\u4f18\u7684\u5ba2\u89c2\u6307\u6807\u3002<\/p>\n<p>Self-attention \u4e0e Cross-attention \u6bd4\u8f83<\/p>\n<p>\u4f5c\u8005\u6307\u51fa\uff0c\u4e3b\u89e3\u7801\u5668\u8f93\u51fa hi \u5df2\u5305\u542b\u5145\u8db3\u4e0a\u4e0b\u6587\u4fe1\u606f\uff0c\u800c\u5b50 token embedding \u7f3a\u4e4f\u4e0a\u4e0b\u6587\u3002\u4ea4\u53c9\u6ce8\u610f\u529b\u5141\u8bb8 hi \u4f5c\u4e3a key\uff0c\u4ece\u800c\u5f62\u6210\u66f4\u76f4\u63a5\u7684\u68af\u5ea6\u4f20\u64ad\u8def\u5f84\uff0c\u56e0\u6b64\u4f18\u4e8e\u81ea\u6ce8\u610f\u529b\u66f4\u65b0 embedding \u7684\u65b9\u5f0f\u3002<\/p>\n<p><a class=\"wp-editor-md-post-content-link\" href=\"https:\/\/tastestars.com\/wp-content\/uploads\/2026\/02\/wp_editor_md_52f26c6d58a91620984a4fe6267c80c2.jpg\"><img decoding=\"async\" src=\"https:\/\/tastestars.com\/wp-content\/uploads\/2026\/02\/wp_editor_md_52f26c6d58a91620984a4fe6267c80c2.jpg\" alt=\"\" \/><\/a><\/p>\n<ol start=\"4\">\n<li>\u5b9e\u9a8c<\/li>\n<\/ol>\n<p>\u7b26\u53f7\u97f3\u4e50\u5b9e\u9a8c\u4f7f\u7528 Pop1k7\u3001Pop909\u3001SOD \u548c LMD clean \u6570\u636e\u96c6\u3002\u79bb\u6563\u97f3\u9891\u5b9e\u9a8c\u4f7f\u7528 MAESTRO \u6570\u636e\u96c6\uff0c\u5e76\u5c06\u97f3\u9891\u7f16\u7801\u4e3a 50Hz \u91c7\u6837\u7387\u7684\u79bb\u6563 token\uff0c\u6bcf\u4e2a token \u5305\u542b 4 \u4e2a codebook\u3002<br \/>\n<a class=\"wp-editor-md-post-content-link\" href=\"https:\/\/tastestars.com\/wp-content\/uploads\/2026\/02\/wp_editor_md_0d4116c5a65be50b5ba5ae7d8fad7009.jpg\"><img decoding=\"async\" src=\"https:\/\/tastestars.com\/wp-content\/uploads\/2026\/02\/wp_editor_md_0d4116c5a65be50b5ba5ae7d8fad7009.jpg\" alt=\"\" \/><\/a><\/p>\n<p>\u6a21\u578b\u89c4\u6a21\u7ea6 4000 \u4e07\u53c2\u6570\uff08\u7b26\u53f7\u97f3\u4e50\uff09\u548c 6200 \u4e07\u53c2\u6570\uff08\u97f3\u9891\uff09\u3002\u4f7f\u7528 AdamW \u4f18\u5316\u5668\u3001cosine \u5b66\u4e60\u7387\u8c03\u5ea6\u3001Flash Attention \u548c\u6df7\u5408\u7cbe\u5ea6\u8bad\u7ec3\u3002<\/p>\n<p>\u7b26\u53f7\u97f3\u4e50\u8bc4\u4f30\u91c7\u7528\u5e73\u5747\u8d1f\u5bf9\u6570\u4f3c\u7136\uff08NLL\uff09\u3002\u4e3a\u516c\u5e73\u6bd4\u8f83\u4e0d\u540c\u7f16\u7801\u65b9\u5f0f\uff0c\u4f5c\u8005\u901a\u8fc7\u79fb\u52a8\u7a97\u53e3\u65b9\u6cd5\u5e76\u5bf9\u6982\u7387\u8fdb\u884c\u6821\u6b63\uff0c\u4f7f\u5176\u5728\u76f8\u4f3c\u4e0a\u4e0b\u6587\u6761\u4ef6\u4e0b\u8fdb\u884c\u6bd4\u8f83\u3002<\/p>\n<p><a class=\"wp-editor-md-post-content-link\" href=\"https:\/\/tastestars.com\/wp-content\/uploads\/2026\/02\/wp_editor_md_718f08b815988106d509f391f7c650d1.jpg\"><img decoding=\"async\" src=\"https:\/\/tastestars.com\/wp-content\/uploads\/2026\/02\/wp_editor_md_718f08b815988106d509f391f7c650d1.jpg\" alt=\"\" \/><\/a><\/p>\n<p>\u7ed3\u679c\u663e\u793a\uff0cNMT \u5728\u6240\u6709\u590d\u5408\u7f16\u7801\u4e0a\u5747\u4f18\u4e8e\u539f\u6a21\u578b\u7ed3\u6784\u3002pitch-first \u7f16\u7801\u5728\u97f3\u9ad8\u9884\u6d4b\u4e0a\u66f4\u4f18\uff0c\u800c metric-first \u5728\u62cd\u4f4d\u9884\u6d4b\u4e0a\u66f4\u4f18\u3002<\/p>\n<p>\u5728\u79bb\u6563\u97f3\u9891\u4efb\u52a1\u4e2d\uff0ccross-attention \u5b50\u89e3\u7801\u5668\u4f18\u4e8e\u81ea\u6ce8\u610f\u529b\u7ed3\u6784\uff0c\u4f46 Embedding Enricher \u672a\u5e26\u6765\u663e\u8457\u6539\u8fdb\u3002\u4f5c\u8005\u63a8\u6d4b\uff0c\u8fd9\u6e90\u4e8e\u7b26\u53f7 token \u4e0e\u97f3\u9891 token \u5728\u8bed\u4e49\u4f9d\u8d56\u7ed3\u6784\u4e0a\u7684\u5dee\u5f02\u3002<\/p>\n<p>\u4e3b\u89c2\u542c\u611f\u6d4b\u8bd5\u7ed3\u679c\u8868\u660e\uff0cNB-PF + NMT \u7684\u6574\u4f53\u8d28\u91cf\u4e0e REMI \u76f8\u5f53\uff0c\u5e76\u4f18\u4e8e Compound Word\u3002<\/p>\n<ol start=\"5\">\n<li>\u7ed3\u8bba<\/li>\n<\/ol>\n<p>\u672c\u6587\u63d0\u51fa\u7684 Nested Music Transformer \u80fd\u591f\u4ee5\u987a\u5e8f\u65b9\u5f0f\u89e3\u7801\u590d\u5408 token\uff0c\u5728\u7b26\u53f7\u97f3\u4e50\u548c\u79bb\u6563\u97f3\u9891\u751f\u6210\u4e2d\u5747\u53d6\u5f97\u7ade\u4e89\u6027\u7ed3\u679c\u3002\u8be5\u67b6\u6784\u901a\u8fc7\u5d4c\u5957 Transformer \u8bbe\u8ba1\uff0c\u540c\u65f6\u89e3\u51b3\u4e86\u5e8f\u5217\u957f\u5ea6\u8fc7\u957f\u548c\u7279\u5f81\u4f9d\u8d56\u5efa\u6a21\u4e0d\u8db3\u7684\u95ee\u9898\uff0c\u5e76\u5728\u4fdd\u6301\u751f\u6210\u8d28\u91cf\u7684\u524d\u63d0\u4e0b\u964d\u4f4e\u4e86\u8bad\u7ec3\u6210\u672c\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u4e00\u7bc7\u5173\u4e8emidi decoder\u89e3\u7801\u5668\u7684\u79d1\u666e\u5de5\u4f5c\uff0chttps:\/\/github.com\/JudeJiwoo\/n [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"jetpack_post_was_ever_published":false,"_jetpack_newsletter_access":"","_jetpack_dont_email_post_to_subs":false,"_jetpack_newsletter_tier_id":0,"_jetpack_memberships_contains_paywalled_content":false,"_jetpack_memberships_contains_paid_content":false,"footnotes":"","jetpack_publicize_message":"","jetpack_publicize_feature_enabled":true,"jetpack_social_post_already_shared":true,"jetpack_social_options":{"image_generator_settings":{"template":"highway","default_image_id":0,"font":"","enabled":false},"version":2}},"categories":[3,52],"tags":[94,106,108],"class_list":["post-774","post","type-post","status-publish","format-standard","hentry","category-studynote","category-52","tag-midi","tag-106","tag-108"],"jetpack_publicize_connections":[],"jetpack_featured_media_url":"","jetpack_sharing_enabled":true,"jetpack_shortlink":"https:\/\/wp.me\/p9Hs6X-cu","_links":{"self":[{"href":"https:\/\/tastestars.com\/index.php\/wp-json\/wp\/v2\/posts\/774","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/tastestars.com\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/tastestars.com\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/tastestars.com\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/tastestars.com\/index.php\/wp-json\/wp\/v2\/comments?post=774"}],"version-history":[{"count":1,"href":"https:\/\/tastestars.com\/index.php\/wp-json\/wp\/v2\/posts\/774\/revisions"}],"predecessor-version":[{"id":775,"href":"https:\/\/tastestars.com\/index.php\/wp-json\/wp\/v2\/posts\/774\/revisions\/775"}],"wp:attachment":[{"href":"https:\/\/tastestars.com\/index.php\/wp-json\/wp\/v2\/media?parent=774"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/tastestars.com\/index.php\/wp-json\/wp\/v2\/categories?post=774"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/tastestars.com\/index.php\/wp-json\/wp\/v2\/tags?post=774"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}