{"id":10021511,"date":"2026-10-28T10:02:23","date_gmt":"2026-10-28T02:02:23","guid":{"rendered":"https:\/\/itmax.cn\/?p=10021511"},"modified":"2026-10-28T10:02:34","modified_gmt":"2026-10-28T02:02:34","slug":"funasr-speech-toolkit-20261028","status":"publish","type":"post","link":"https:\/\/itmax.cn\/?p=10021511","title":{"rendered":"\u4e09\u4e2a\u5c0f\u65f6\u7684\u4f1a\u6211\u53ea\u60f3\u8981\u4e00\u4efd\u80fd\u7528\u7684\u6587\u5b57\uff5cFunASR \u88c5\u5b8c\uff0c\u6211\u5361\u5728\u65ad\u53e5\u4e0a"},"content":{"rendered":"<figure style=\"margin:0 0 32px;text-align:center;\"><img decoding=\"async\" src=\"https:\/\/itmax.cn\/wp-content\/uploads\/2026\/10\/funasr-cover.png\" alt=\"FunASR \u9879\u76ee\u5c01\u9762\uff1a\u4e2d\u6587\u8bed\u97f3\u8f6c\u5199\u5de5\u5177\u5305\uff0c\u56db\u9897\u5c0f\u6a21\u578b\u4e32\u6210\u4e00\u6761\u6d41\u6c34\u7ebf\" style=\"width:100%;height:auto;max-width:960px;border-radius:8px;display:block;margin:0 auto;\" \/><figcaption style=\"margin-top:12px;font-size:14px;color:#868EAA;line-height:1.6;\">FunASR \u5728 GitHub \u4e0a\u7684\u5f53\u524d\u72b6\u6001\uff1a20,845 star\u3001v1.4.16\u3001\u8fd1\u56db\u5468 43 \u6b21\u63d0\u4ea4<\/figcaption><\/figure>\n<p>\u4e0a\u5468\u56db\u4e0b\u5348\u8fde\u7740\u5f00\u4e86\u4e09\u4e2a\u4f1a\uff0c\u5f55\u97f3\u52a0\u8d77\u6765\u56db\u5c0f\u65f6\u4e8c\u5341\u5206\u949f\u3002\u6211\u7684\u4e60\u60ef\u662f\u5f53\u5929\u665a\u4e0a\u628a\u5f55\u97f3\u8fc7\u4e00\u904d\uff0c\u6458\u51fa\u51b3\u8bae\u548c\u5f85\u529e\uff0c\u7b2c\u4e8c\u5929\u65e9\u4f1a\u76f4\u63a5\u7528\u3002\u8fd9\u4e2a\u4e60\u60ef\u7ef4\u6301\u4e86\u5927\u534a\u5e74\uff0c\u9760\u7684\u662f\u4e00\u5957\u81ea\u5df1\u62fc\u7684\u672c\u5730\u8f6c\u5199\uff1a\u5148\u62ff whisper.cpp \u8dd1\u4e00\u904d\uff0c\u518d\u7528\u811a\u672c\u628a\u7ed3\u679c\u5207\u6bb5\uff0c\u6700\u540e\u4eba\u5de5\u6821\u5bf9\u3002<\/p>\n<p>\u90a3\u5929\u665a\u4e0a\u6821\u5bf9\u5230\u7b2c\u4e8c\u4e2a\u5c0f\u65f6\uff0c\u6211\u5173\u6389\u4e86\u7a97\u53e3\u3002<\/p>\n<p>\u95ee\u9898\u4e0d\u662f\u5b83\u542c\u4e0d\u61c2\uff0c\u662f\u5b83\u300c\u542c\u4e0d\u61c2\u7684\u5730\u65b9\u7279\u522b\u7a33\u5b9a\u5730\u9519\u5728\u540c\u4e00\u4e2a\u65b9\u5411\u300d\u3002\u6211\u4eec\u5185\u90e8\u628a\u300c\u5408\u540c\u5c65\u7ea6\u300d\u5ff5\u5feb\u7684\u65f6\u5019\uff0c\u5b83\u7ed9\u51fa\u300c\u5408\u540c\u65c5\u7ea6\u300d\uff1b\u540c\u4e8b\u8bf4\u300c\u8fd9\u4e2a\u5b63\u5ea6\u7684 kpi \u53ea\u6709\u534e\u4e1c\u8fbe\u6807\u300d\uff0c\u5b83\u628a kpi \u5199\u6210\u300cKP I\u300d\uff0c\u4e2d\u95f4\u585e\u4e2a\u7a7a\u683c\uff0c\u6211\u6309\u7a7a\u683c\u5207\u8bcd\u7684\u540e\u5904\u7406\u6574\u53e5\u90fd\u5bf9\u4e0d\u9f50\uff1b\u8fd8\u6709\u4e00\u6bb5\u4e2d\u82f1\u6df7\u7740\u8bb2\uff0c\u300c\u8fd9\u4e2a case \u6211\u4eec escalate \u5230\u6cd5\u52a1\u4e86\u300d\u8f6c\u6210\u4e86\u300c\u8fd9\u4e2a K \u6b7b\u6211\u4eec S \u5f00\u4e86\u5230\u6cd5\u52a1\u4e86\u300d\u3002<\/p>\n<p>\u8fd9\u7c7b\u9519\u8bef\u6700\u70e6\u7684\u662f\u5b83\u4e0d\u662f\u968f\u673a\u566a\u58f0\uff0c\u662f\u6211\u6bcf\u7bc7\u90fd\u8981\u624b\u5de5\u6539\u4e00\u904d\u7684\u9ad8\u9891\u8bcd\u3002\u4e00\u767e\u53e5\u9519\u516b\u53e5\uff0c\u542c\u7740\u8fd8\u6709\u4e5d\u6210\u51c6\u786e\u7387\uff0c\u4f46\u843d\u5230\u300c\u62ff\u5b83\u5f53\u7eaa\u8981\u5e95\u7a3f\u300d\u8fd9\u4ef6\u4e8b\u4e0a\uff0c\u4e5d\u6210\u7b49\u4e8e\u96f6\u2014\u2014\u4f60\u5fc5\u987b\u9010\u53e5\u8bfb\uff0c\u624d\u6562\u786e\u8ba4\u54ea\u53e5\u80fd\u7528\u3002<\/p>\n<p>\u6211\u8bd5\u8fc7\u4e09\u6761\u8def\u3002\u6362\u66f4\u5927\u7684 whisper \u6a21\u578b\uff0c\u540c\u97f3\u5b57\u597d\u4e00\u70b9\uff0c\u901f\u5ea6\u6389\u5230 0.4 \u500d\u5b9e\u65f6\uff0c\u4e00\u53f0 M2 \u8dd1\u56db\u5c0f\u65f6\u5f55\u97f3\u8981\u5341\u5c0f\u65f6\u3002\u8d70\u4e91\u4e0a\u7684\u63a5\u53e3\uff0c\u6309\u5206\u949f\u8ba1\u8d39\u4e0d\u8d35\uff0c\u53ef\u4f1a\u8bae\u5185\u5bb9\u91cc\u6709\u5ba2\u6237\u540d\u5b57\u3001\u6709\u6ca1\u516c\u5f00\u7684\u6392\u671f\uff0c\u6211\u4e0d\u653e\u5fc3\u3002\u7b2c\u4e09\u6761\u662f\u81ea\u5df1\u8bad\uff0c\u770b\u4e86\u4e00\u773c\u6807\u6ce8\u6210\u672c\u5c31\u653e\u5f03\u4e86\u3002<\/p>\n<p>\u6240\u4ee5\u5f53\u6211\u7ffb\u5230 FunASR \u90a3\u53e5\u300c\u4e2d\u6587 CER \u53ea\u6709 whisper \u7cfb\u4e09\u5206\u4e4b\u4e00\u300d\u7684\u8bf4\u6cd5\u65f6\uff0c\u7b2c\u4e00\u53cd\u5e94\u662f\u4e0d\u4fe1\u3002\u8fbe\u6469\u9662\u8fd9\u5957\u4e1c\u897f\u6211\u65e9\u6709\u8033\u95fb\uff0c2022 \u5e74\u5e95\u5c31\u5f00\u6e90\u4e86\uff0c\u4e00\u76f4\u4ee5\u4e3a\u662f\u7ed9\u5b66\u672f\u754c\u8dd1 benchmark \u7684\uff0c\u5de5\u7a0b\u4e0a\u4e0d\u597d\u7528\u3002\u8fd9\u6b21\u6211\u51b3\u5b9a\u8ba4\u771f\u8fc7\u4e00\u904d\uff0c\u7406\u7531\u662f\u5b83\u6700\u8fd1\u7684\u66f4\u65b0\u8282\u594f\u53d8\u4e86\uff1a\u52a0\u4e86 OpenAI \u517c\u5bb9\u7684 HTTP \u63a5\u53e3\u3001\u52a0\u4e86 llama.cpp \u7684\u5355\u6587\u4ef6\u8fd0\u884c\u65f6\u3001\u52a0\u4e86\u547d\u4ee4\u884c\u5de5\u5177\uff0c\u8fd9\u4e9b\u4e1c\u897f\u4e0d\u662f\u8bba\u6587\u9879\u76ee\u4f1a\u505a\u7684\u4e8b\u3002<\/p>\n<h2 id=\"_1\">\u5b83\u5176\u5b9e\u662f\u56db\u9897\u5c0f\u6a21\u578b\u4e32\u6210\u7684\u4e00\u6761\u6d41\u6c34\u7ebf<\/h2>\n<p>\u6211\u4e00\u5f00\u59cb\u4ee5\u4e3a <code>pip install funasr<\/code> \u88c5\u7684\u662f\u300c\u4e00\u4e2a\u6a21\u578b\u300d\u3002\u8bfb\u5b8c README \u7684\u6a21\u578b\u6e05\u5355\u624d\u660e\u767d\uff0c\u5b83\u5356\u7684\u662f\u4e00\u5957<strong>\u7ec4\u4ef6\u62fc\u88c5<\/strong>\u7684\u601d\u8def\uff0c\u6a21\u578b\u53ef\u4ee5\u81ea\u7531\u6362\u3002<\/p>\n<p>\u9aa8\u67b6\u4e0a\u6709\u56db\u4e2a\u4f4d\u7f6e\uff1a<\/p>\n<table>\n<thead>\n<tr>\n<th>\u4f4d\u7f6e<\/th>\n<th>\u5e72\u4ec0\u4e48<\/th>\n<th>\u5e38\u7528\u578b\u53f7<\/th>\n<th>\u53c2\u6570\u91cf<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>VAD<\/td>\n<td>\u627e\u51fa\u54ea\u51e0\u6bb5\u6709\u4eba\u8bf4\u8bdd<\/td>\n<td><code>fsmn-vad<\/code><\/td>\n<td>0.4M<\/td>\n<\/tr>\n<tr>\n<td>ASR<\/td>\n<td>\u628a\u8bed\u97f3\u8f6c\u6210\u5b57<\/td>\n<td><code>SenseVoiceSmall<\/code> \/ <code>Paraformer-zh<\/code><\/td>\n<td>234M \/ 220M<\/td>\n<\/tr>\n<tr>\n<td>PUNC<\/td>\n<td>\u65ad\u53e5\u52a0\u6807\u70b9<\/td>\n<td><code>ct-punc<\/code><\/td>\n<td>290M<\/td>\n<\/tr>\n<tr>\n<td>SPK<\/td>\n<td>\u8c01\u5728\u8bf4<\/td>\n<td><code>cam++<\/code><\/td>\n<td>7.2M<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p>\u6ce8\u610f VAD \u90a3\u4e2a 0.4M \u548c\u8bf4\u8bdd\u4eba\u90a3\u4e2a 7.2M\u2014\u2014\u6574\u6761\u94fe\u4e0a\u6700\u4e0d\u503c\u94b1\u7684\u4e24\u4e2a\uff0c\u4f46\u7f3a\u4e86\u5b83\u4eec\u8f93\u51fa\u5c31\u6ca1\u6cd5\u7528\u3002\u6ca1\u6709 VAD\uff0c\u957f\u97f3\u9891\u4f1a\u4e00\u6b21\u6027\u5582\u8fdb\u6a21\u578b\uff0c\u8d85\u8fc7\u4e09\u5341\u79d2\u5c31\u5f00\u59cb\u4e22\u5185\u5bb9\uff1b\u6ca1\u6709 SPK\uff0c\u591a\u4eba\u4f1a\u8bae\u8f6c\u51fa\u6765\u662f\u4e00\u5768\u6ca1\u5f52\u5c5e\u7684\u6587\u5b57\u3002<\/p>\n<p>\u7ec4\u4ef6\u5316\u6700\u76f4\u63a5\u7684\u7528\u6cd5\u662f\u4e00\u884c\u88c5\u9f50\uff1a<\/p>\n<pre><code class=\"language-python\">from funasr import AutoModel\n\nmodel = AutoModel(\n    model=&quot;paraformer-zh&quot;,\n    vad_model=&quot;fsmn-vad&quot;,\n    punc_model=&quot;ct-punc&quot;,\n    spk_model=&quot;cam++&quot;,\n    device=&quot;cpu&quot;,\n)\nresult = model.generate(input=&quot;meeting.wav&quot;, batch_size_s=300)\n<\/code><\/pre>\n<p>\u8fd9\u6761\u94fe\u8dd1\u5b8c\uff0c<code>result[0][\"sentence_info\"]<\/code> \u91cc\u662f\u4e00\u6bb5\u4e00\u6bb5\u7684\u7ed3\u6784\u5316\u7ed3\u679c\uff0c\u6bcf\u6bb5\u5e26 <code>start<\/code>\u3001<code>end<\/code>\u3001<code>spk<\/code>\u3001<code>sentence<\/code> \u56db\u4e2a\u5b57\u6bb5\u3002\u8fd9\u5c31\u662f\u6211\u8981\u7684\u4e1c\u897f\uff1a\u65f6\u95f4\u6233\u80fd\u5b9a\u4f4d\u5230\u5f55\u97f3\u7b2c\u51e0\u5206\u51e0\u79d2\uff0c<code>spk<\/code> \u80fd\u533a\u5206\u662f\u51e0\u4e2a\u4eba\u5728\u8bf4\u8bdd\u3002<\/p>\n<p>\u7b2c\u4e8c\u4ef6\u6211\u786e\u8ba4\u80fd\u7528\u7684\u662f\u547d\u4ee4\u884c\u3002\u8fd9\u4e2a\u4ed3\u5e93 1.4 \u7248\u4e4b\u540e\u628a CLI \u505a\u6210\u4e86\u6b63\u7ecf\u5165\u53e3\uff0c<code>setup.py<\/code> \u91cc\u6302\u4e86\u4e94\u4e2a console script\uff1a<code>funasr<\/code>\u3001<code>funasr-server<\/code>\u3001<code>funasr-realtime-server<\/code>\u3001<code>funasr-export<\/code>\u3001<code>funasr-train<\/code>\u3002\u5bf9\u6211\u8fd9\u79cd\u300c\u4e00\u4e2a\u6587\u4ef6\u5939\u91cc\u4e8c\u5341\u4e2a wav\u300d\u7684\u573a\u666f\uff0cCLI \u6bd4\u5199\u811a\u672c\u5feb\uff1a<\/p>\n<pre><code class=\"language-bash\"># \u6700\u7b80\uff1a\u8f6c\u4e00\u4e2a\u6587\u4ef6\uff0c\u76f4\u63a5\u6253\u5370\u6587\u5b57\nfunasr meeting.wav\n\n# \u5e26\u8bf4\u8bdd\u4eba\u548c\u65f6\u95f4\u6233\uff0c\u8f93\u51fa JSON\nfunasr meeting.wav --spk --timestamps -f json\n\n# \u6279\u91cf\u51fa SRT \u5b57\u5e55\uff0c\u5199\u5230 .\/subs\nfunasr *.wav -f srt -o .\/subs\n\n# \u6307\u5b9a\u6a21\u578b\u548c\u8bed\u79cd\nfunasr meeting.wav -m paraformer -l zh\n<\/code><\/pre>\n<p>\u7b2c\u4e09\u4ef6\u662f\u628a\u5b83\u53d8\u6210\u672c\u5730 HTTP \u670d\u52a1\uff0c\u63a5\u53e3\u6309 OpenAI \u7684\u97f3\u9891\u8f6c\u5199\u683c\u5f0f\u6765\u3002\u6211\u539f\u672c\u89c9\u5f97\u8fd9\u5c31\u662f\u300c\u8e6d\u683c\u5f0f\u300d\uff0c\u540e\u6765\u60f3\u660e\u767d\u5b83\u7684\u4ef7\u503c\uff1a\u6211\u624b\u4e0a\u90a3\u4e9b\u7b14\u8bb0\u63d2\u4ef6\u3001\u81ea\u52a8\u5316\u6d41\u7a0b\u90fd\u53ea\u8ba4\u4e00\u4e2a <code>base_url<\/code>\uff0c\u672c\u5730\u670d\u52a1\u4e00\u8d77\u6765\uff0c\u5b83\u4eec\u76f4\u63a5\u63a5\u4e0a\uff0c\u4e24\u8fb9\u4ee3\u7801\u90fd\u4e0d\u7528\u6539\u3002<\/p>\n<pre><code class=\"language-bash\">funasr-server --host 127.0.0.1 --port 8000 --model sensevoice --device cpu\n<\/code><\/pre>\n<p>\u8d77\u597d\u4e4b\u540e <code>POST \/v1\/audio\/transcriptions<\/code>\uff0c<code>-F model=sensevoice<\/code>\uff0c\u8fd4\u56de\u683c\u5f0f\u548c\u4e91\u7aef\u90a3\u5957\u4e00\u81f4\u3002\u5b98\u65b9\u6587\u6863\u91cc\u4e13\u95e8\u5199\u4e86\u4e00\u53e5\u63d0\u9192\uff1a\u8fd9\u4e2a\u670d\u52a1\u9ed8\u8ba4\u4e0d\u5e26\u9274\u6743\uff0c\u53ea\u8be5\u76d1\u542c\u5728 loopback \u4e0a\u3002\u8fd9\u8bdd\u662f\u5bf9\u7684\uff0c\u6211\u89c1\u8fc7\u592a\u591a\u4eba\u628a\u8fd9\u79cd\u670d\u52a1\u5f00\u5230 <code>0.0.0.0<\/code> \u7136\u540e\u5fd8\u4e86\u3002<\/p>\n<p>\u8fd8\u6709\u4e00\u5757\u6211\u4e0d\u592a\u4e70\u8d26\uff1a<strong>\u60c5\u611f\u548c\u97f3\u9891\u4e8b\u4ef6\u6807\u7b7e<\/strong>\u3002SenseVoice \u7684\u8f93\u51fa\u4f1a\u5e26 <code>&lt;|HAPPY|&gt;<\/code>\u3001<code>&lt;|BGM|&gt;<\/code> \u8fd9\u7c7b\u6807\u8bb0\uff0c<code>rich_transcription_postprocess<\/code> \u8d1f\u8d23\u5265\u6389\u3002\u542c\u7740\u80fd\u7edf\u8ba1\u300c\u8fd9\u573a\u4f1a\u8d1f\u9762\u60c5\u7eea\u5360\u6bd4\u300d\uff0c\u5b9e\u9645\u7528\u9014\u6709\u9650\u2014\u2014\u4e00\u6bb5\u8bdd\u88ab VAD \u5207\u6210\u5341\u51e0\u6bb5\uff0c\u6bcf\u6bb5\u4e00\u4e2a\u60c5\u7eea\u6807\u7b7e\uff0c\u4f60\u5f97\u5230\u7684\u662f\u5341\u51e0\u4e2a\u4eba\u5de5\u65e0\u6cd5\u9a8c\u8bc1\u7684\u5224\u65ad\u3002\u6211\u7684\u5224\u65ad\u662f\uff0c\u5b83\u5728\u5ba2\u670d\u8d28\u68c0\u8fd9\u79cd\u77ed\u53e5\u3001\u5355\u4eba\u3001\u6709\u660e\u786e\u4e1a\u52a1\u89c4\u5219\u7684\u573a\u666f\u91cc\u53ef\u80fd\u6709\u7528\uff0c\u5728\u957f\u4f1a\u8bae\u91cc\u662f\u7eaf\u566a\u58f0\u3002<\/p>\n<p>\u540c\u6837\u504f\u5671\u5934\u7684\u8fd8\u6709\u4e00\u4e2a\uff1a\u9996\u9875\u628a\u300cvLLM \u52a0\u901f\u300d\u6392\u5f97\u5f88\u663e\u773c\uff0c\u4f46\u5b98\u65b9\u90e8\u7f72\u6587\u6863\u81ea\u5df1\u627f\u8ba4\uff0cvLLM \u90a3\u6761\u8def\u53ea\u9002\u7528\u4e8e Fun-ASR-Nano \u8fd9\u79cd\u81ea\u56de\u5f52\u6a21\u578b\uff0c\u5bf9 Paraformer \u8fd9\u79cd\u975e\u81ea\u56de\u5f52\u7684<strong>\u4e0d\u9002\u7528<\/strong>\u3002\u60f3\u52a0\u901f\u4e2d\u6587\u751f\u4ea7\u8def\u5f84\u4e0a\u6700\u6210\u719f\u7684\u6a21\u578b\uff0c\u8fd9\u4e2a\u65b9\u6848\u8ddf\u4f60\u65e0\u5173\u3002<\/p>\n<h2 id=\"_2\">\u975e\u81ea\u56de\u5f52\u7701\u4e0b\u7684\u65f6\u95f4\uff0c\u4ece\u65f6\u95f4\u6233\u91cc\u6263\u8d70\u4e86<\/h2>\n<p>FunASR \u4e2d\u6587\u7ebf\u7684\u6838\u5fc3\u53d6\u820d\uff0c\u4e00\u53e5\u8bdd\u80fd\u8bf4\u5b8c\uff1a<strong>\u7528\u975e\u81ea\u56de\u5f52\u6362\u901f\u5ea6<\/strong>\u3002<\/p>\n<p>whisper \u662f\u81ea\u56de\u5f52\u7684\uff0c\u4e00\u4e2a\u5b57\u4e00\u4e2a\u5b57\u5f80\u5916\u5410\uff0c\u6bcf\u4e2a\u5b57\u90fd\u4f9d\u8d56\u524d\u9762\u5df2\u7ecf\u5410\u51fa\u6765\u7684\u5b57\u3002\u597d\u5904\u662f\u4e0a\u4e0b\u6587\u8fde\u8d2f\u3001\u80fd\u76f4\u63a5\u8f93\u51fa\u7ffb\u8bd1\uff1b\u574f\u5904\u662f\u6162\uff0c\u800c\u4e14\u957f\u97f3\u9891\u5bb9\u6613\u300c\u590d\u8bfb\u300d\u2014\u2014\u540c\u4e00\u4e2a\u53e5\u5b50\u5faa\u73af\u8f93\u51fa\uff0c\u8dd1 whisper \u7684\u4eba\u5e94\u8be5\u90fd\u9047\u5230\u8fc7\u3002<\/p>\n<p>Paraformer \u548c SenseVoice \u8d70\u7684\u662f\u53e6\u4e00\u6761\u8def\uff1a\u4e00\u6b21\u524d\u5411\uff0c\u76f4\u63a5\u5410\u51fa\u6574\u6bb5\u5b57\u5e8f\u5217\uff0c\u4e0d\u505a\u9010\u5b57\u89e3\u7801\u3002\u8fd9\u5c31\u662f\u901f\u5ea6\u5dee\u7684\u6839\u672c\u539f\u56e0\uff0c\u4e0d\u662f\u5de5\u7a0b\u4f18\u5316\u505a\u5f97\u597d\uff0c\u662f\u7b97\u6cd5\u7ed3\u6784\u4e0a\u5c11\u4e86\u4e00\u4e2a\u5faa\u73af\u3002\u5b98\u65b9\u7ad9\u70b9\u7ed9\u7684\u6570\u5b57\u662f\u7eaf CPU \u4e0a\u7ea6\u4e8c\u5341\u500d\u5b9e\u65f6\uff0c\u6211\u81ea\u5df1\u5728 M2 \u4e0a\u6d4b\u77ed\u97f3\u9891\u6ca1\u90a3\u4e48\u597d\u770b\uff08\u4e0b\u9762\u6709\u5177\u4f53\u6570\uff09\uff0c\u4f46\u5dee\u4e00\u4e2a\u91cf\u7ea7\u7684\u90a3\u4e2a\u65b9\u5411\u662f\u5bf9\u7684\u3002<\/p>\n<p>\u4ee3\u4ef7\u6709\u4e09\u4e2a\uff0c\u800c\u4e14\u90fd\u5f88\u5b9e\u5728\u3002<\/p>\n<p><strong>\u7b2c\u4e00\uff0c\u65f6\u95f4\u6233\u7684\u7c92\u5ea6\u662f\u7c97\u7684\uff0c\u800c\u4e14\u56fa\u5b9a\u7c97\u3002<\/strong> \u6211\u8fd9\u6b21\u52a0\u4e86 <code>--timestamps<\/code>\uff0c\u62ff\u5230\u7684\u662f\u9010\u5b57\u65f6\u95f4\u6bb5\uff1a\u4e00\u53e5\u8bdd\u5341\u4e5d\u4e2a\u5b57\uff0c\u5341\u4e5d\u4e2a\u65f6\u95f4\u6bb5\uff0c\u6bcf\u6bb5\u5bbd\u5ea6<strong>\u6b63\u597d 60 \u6beb\u79d2<\/strong>\uff0c\u4e00\u4e2a\u4e0d\u5dee\u3002\u8fd9\u4e0d\u662f\u5de7\u5408\uff0c\u662f\u8fd9\u7c7b\u6a21\u578b\u7684\u8f93\u51fa\u7c92\u5ea6\u2014\u2014token \u4e00\u683c 60ms\u3002\u6240\u4ee5\u6211\u77e5\u9053\u6bcf\u4e2a\u5b57\u5927\u6982\u843d\u5728\u7b2c\u51e0\u79d2\uff0c\u4f46\u62ff\u4e0d\u5230\u300c\u4ece 62 \u6beb\u79d2\u5f00\u59cb\u300d\u8fd9\u79cd\u7cbe\u5ea6\u3002\u505a\u5b57\u5e55\u591f\u7528\uff0c\u505a\u300c\u70b9\u67d0\u4e2a\u8bcd\u8df3\u56de\u5f55\u97f3\u90a3\u4e00\u5e27\u300d\u5c31\u7cd9\u4e86\u3002\u8981\u66f4\u7ec6\u53ea\u80fd\u6362 Paraformer \u7684 CTC \u8f93\u51fa\uff0c\u6216\u8005\u4e0a Fun-ASR-Nano\u3002<\/p>\n<p><strong>\u7b2c\u4e8c\uff0c\u8de8\u6bb5\u7684\u4e0a\u4e0b\u6587\u4e22\u4e86\u3002<\/strong> VAD \u628a\u97f3\u9891\u5207\u6210\u4e00\u6bb5\u4e00\u6bb5\uff0c\u6bcf\u6bb5\u72ec\u7acb\u9001\u8fdb ASR\u3002\u6bb5\u4e0e\u6bb5\u4e4b\u95f4\u6ca1\u6709\u4e0a\u4e0b\u6587\u4f20\u9012\uff0c\u6240\u4ee5\u4e00\u53e5\u8bdd\u88ab VAD \u5207\u5728\u4e24\u6bb5\u91cc\u7684\u65f6\u5019\uff0c\u524d\u534a\u6bb5\u7ed3\u5c3e\u548c\u540e\u534a\u6bb5\u5f00\u5934\u5404\u8f6c\u4e00\u6b21\uff0c\u5f88\u5bb9\u6613\u5404\u9519\u4e00\u534a\u3002\u8fd9\u4e5f\u662f\u4e3a\u4ec0\u4e48\u5f88\u591a\u4eba\u5728 issue \u91cc\u62b1\u6028\u300c\u52a0\u4e86\u70ed\u8bcd\u4e4b\u540e\u53e5\u5b50\u53d8\u77ed\u4e86\u300d\u2014\u2014\u70ed\u8bcd\u66ff\u6362\u662f\u6309\u6bb5\u505a\u7684\uff0c\u6bb5\u8fb9\u754c\u4e00\u53d8\uff0c\u66ff\u6362\u7684\u843d\u70b9\u5c31\u8ddf\u7740\u53d8\u3002<\/p>\n<p><strong>\u7b2c\u4e09\uff0c\u4e5f\u662f\u6211\u8ba4\u4e3a\u6700\u9700\u8981\u63d0\u524d\u60f3\u6e05\u695a\u7684\uff1a\u5b83\u7684\u300c\u51c6\u300d\u662f\u5efa\u7acb\u5728\u6807\u70b9\u6a21\u578b\u548c\u5206\u6bb5\u7b56\u7565\u4e0a\u7684\uff0c\u4e0d\u662f\u5efa\u7acb\u5728\u8bed\u8a00\u6a21\u578b\u4e0a\u7684\u3002<\/strong> whisper \u5185\u90e8\u6709\u8bed\u8a00\u6a21\u578b\u5728\u5e2e\u7740\u7ea0\u6b63\uff0c\u6240\u4ee5\u4f60\u5ff5\u9519\u4e00\u4e2a\u5b57\u5b83\u53ef\u80fd\u66ff\u4f60\u731c\u5bf9\uff1bFunASR \u8fd9\u6761\u94fe\u4e0a\uff0c\u731c\u5bf9\u7684\u8d23\u4efb\u5728 <code>ct-punc<\/code> \u548c\u540e\u7eed\u66ff\u6362\u89c4\u5219\u3002\u8fd9\u610f\u5473\u7740\u4e24\u4ef6\u4e8b\uff1a<code>ct-punc<\/code> \u7684\u7248\u672c\u4f1a\u76f4\u63a5\u5f71\u54cd\u4f60\u7684\u65ad\u53e5\u8d28\u91cf\uff1b\u70ed\u8bcd\u529f\u80fd\u662f\u53cc\u5203\u5251\u3002<\/p>\n<p>\u8fd9\u4e09\u6761\u4ee3\u4ef7\u4ec0\u4e48\u65f6\u5019\u4e0d\u53ef\u63a5\u53d7\uff1f<strong>\u8981\u8bcd\u7ea7\u5bf9\u9f50\u3001\u8981\u7ffb\u8bd1\u3001\u8981\u5c0f\u8bed\u79cd<\/strong>\uff0c\u8fd9\u4e09\u79cd\u522b\u7528\u5b83\u3002\u7ffb\u8bd1\u80fd\u529b\u57fa\u672c\u6ca1\u6709\uff0c\u5b98\u65b9\u6e05\u5355\u91cc whisper-large-v3 \u662f\u88ab\u5f53\u4f5c\u300c\u4e00\u4e2a\u53ef\u9009\u6a21\u578b\u300d\u6536\u8fdb\u6765\u7684\uff0c\u4e0d\u662f\u5bf9\u6807\u5bf9\u8c61\uff1b\u5c0f\u8bed\u79cd\u90a3\u8fb9\u6700\u65b0\u7684 MLT-Nano \u8986\u76d6 31 \u79cd\u8bed\u8a00\uff0c\u4f46\u4e2d\u6587\u5708\u4e4b\u5916\u7684\u8868\u73b0\uff0c\u5b98\u65b9\u81ea\u5df1\u4e5f\u6ca1\u7ed9\u53ef\u6bd4\u6570\u636e\u3002<\/p>\n<h2 id=\"torch\">\u88c5\u4e4b\u524d\u5148\u628a torch \u8fd9\u5173\u8fc7\u4e86<\/h2>\n<p>\u5b98\u65b9 README \u7ed9\u7684\u5b89\u88c5\u6b65\u9aa4\u5c31\u4e24\u884c\uff0c\u6211\u7167\u7740\u8d70\uff0c\u4e2d\u95f4\u5728\u4e0b\u8f7d\u4e0a\u683d\u4e86\u4e00\u6b21\uff1a<\/p>\n<pre><code class=\"language-bash\"># 1. \u5148\u88c5\u5339\u914d\u7684 torch \u548c torchaudio\uff08\u8fd9\u4e00\u6b65\u51b3\u5b9a\u540e\u9762\u4e00\u5207\uff09\npip install torch torchaudio\n\n# 2. \u786e\u8ba4\u8bbe\u5907\u80fd\u88ab\u8bc6\u522b\u5230\npython -c &quot;import torch; print(torch.cuda.is_available())&quot;\n\n# 3. \u518d\u88c5 funasr\npip install funasr\n<\/code><\/pre>\n<p>\u6211\u7b2c\u4e00\u6b21\u662f\u76f4\u63a5\u8fde PyPI \u88c5\u7684\uff0c<code>pip install torch torchaudio<\/code> \u8dd1\u4e86\u4e8c\u5341\u4e94\u5206\u949f\uff0c\u8fdb\u7a0b\u8fd8\u5728\u3001\u7f13\u5b58\u4e0d\u6da8\u3001CPU \u5360\u7528\u96f6\u70b9\u51e0\u2014\u2014\u5178\u578b\u7684\u8fde\u63a5\u6302\u5728\u90a3\u91cc\u4e0d\u52a8\u3002\u6740\u6389\u6362\u6210\u6e05\u534e\u6e90\uff0c<strong>\u4e09\u5341\u516d\u79d2\u88c5\u5b8c<\/strong>\uff0ctorch 2.14.1\u3001torchaudio 2.11.0\u3002\u63a5\u7740\u88c5 funasr 1.4.16\uff0c\u4e00\u5206\u96f6\u516b\u79d2\uff0c\u987a\u5e26\u88c5\u4e86\u4e03\u5341\u591a\u4e2a\u4f9d\u8d56\u3002\u56fd\u5185\u7f51\u7edc\u73af\u5883\u4e0b\u8fd9\u4e00\u6b65\u522b\u786c\u7b49\uff0c\u6362\u6e90\u662f\u6700\u7701\u65f6\u95f4\u7684\u52a8\u4f5c\u3002<\/p>\n<p>\u7b2c\u4e8c\u6b65\u90a3\u4e2a\u68c0\u67e5\uff0c\u5728 Mac \u4e0a\u5fc5\u7136\u662f <code>False<\/code>\uff0c\u6240\u4ee5\u6211\u6ca1\u5f97\u9009\uff0c\u53ea\u80fd\u8d70 CPU\u3002\u8fd9\u4e00\u6b65\u503c\u5f97\u5199\u4e0b\u6765\uff0c\u56e0\u4e3a <code>AutoModel<\/code> \u5bf9\u4e0d\u53ef\u7528\u7684 CUDA \u662f<strong>\u9759\u9ed8\u56de\u9000<\/strong>\u7684\uff1a\u5199\u4e86 <code>device=\"cuda\"<\/code> \u4e0d\u62a5\u9519\uff0c\u5b89\u5b89\u9759\u9759\u5728 CPU \u4e0a\u8dd1\u5b8c\uff0c\u6162\u4e09\u500d\u4f60\u8fd8\u4ee5\u4e3a\u6a21\u578b\u5c31\u8fd9\u6c34\u5e73\u3002\u5b98\u65b9\u6392\u969c\u6587\u6863\u4e3a\u6b64\u4e13\u95e8\u7ed9\u4e86\u4e00\u6bb5\u811a\u672c\uff0c\u9010\u4e2a\u6253\u5370 ASR\u3001VAD\u3001PUNC\u3001SPK \u56db\u4e2a\u6a21\u578b\u53c2\u6570\u5b9e\u9645\u6240\u5728\u7684\u8bbe\u5907\uff0c\u6ce8\u91ca\u5199\u5f97\u5f88\u76f4\u767d\u2014\u2014\u300c\u8bf7\u6c42\u7684\u8bbe\u5907\u4e0d\u7b49\u4e8e\u5b9e\u9645\u7684\u653e\u7f6e\u300d\u3002\u8fd9\u6bb5\u6211\u6ca1\u5728 CUDA \u673a\u5668\u4e0a\u9a8c\u8fc7\uff0c\u4f46\u5b83\u5b58\u5728\u672c\u8eab\u5c31\u8bf4\u660e\u8e29\u7684\u4eba\u591f\u591a\u3002<\/p>\n<p><code>pip check<\/code> \u4e5f\u503c\u5f97\u8dd1\u4e00\u6b21\u3002\u4f9d\u8d56\u5217\u8868\u6bd4\u770b\u4e0a\u53bb\u957f\uff1a<code>modelscope<\/code>\u3001<code>huggingface_hub<\/code>\u3001<code>transformers<\/code>\u3001<code>hydra-core<\/code>\u3001<code>omegaconf<\/code>\u3001<code>umap_learn<\/code>\u3001<code>oss2<\/code>\u3001<code>kaldiio<\/code>\u3001<code>librosa<\/code>\uff0c<code>umap_learn<\/code> \u662f\u4e3a\u4e86\u8bf4\u8bdd\u4eba\u805a\u7c7b\uff0c<code>oss2<\/code> \u662f\u5bf9\u8c61\u5b58\u50a8 SDK\uff0c<code>hydra-core<\/code> \u662f\u4e3a\u4e86\u8bad\u7ec3\u914d\u7f6e\u3002\u53ea\u60f3\u8f6c\u4e2a\u97f3\uff0c\u5374\u987a\u5e26\u88c5\u4e86\u534a\u4e2a\u8bad\u7ec3\u6846\u67b6\u3002<\/p>\n<p><strong>\u786c\u4ef6\u95e8\u69db\u8fd9\u4ef6\u4e8b\uff0c\u5b98\u65b9\u6ca1\u6709\u7ed9\u6574\u673a\u914d\u7f6e\u8868\u3002<\/strong> \u6211\u7ffb\u4e86 README\u3001\u90e8\u7f72\u77e9\u9635\u3001\u6392\u969c\u6587\u6863\uff0c\u53ea\u627e\u5230\u4e00\u53e5\u300cPython \u2265 3.8\u300d\uff0c\u5185\u5b58\u548c\u663e\u5b58\u7684\u8981\u6c42\u4e00\u4e2a\u5b57\u6ca1\u6709\u3002\u6240\u4ee5\u4e0b\u9762\u4e24\u6863\u662f\u6211\u7684\u63a8\u7b97\uff0c\u4e0d\u662f\u5b98\u65b9\u6570\u5b57\uff0c\u63a8\u7b97\u4f9d\u636e\u6211\u5199\u6e05\u695a\uff1a<\/p>\n<table>\n<thead>\n<tr>\n<th>\u6863\u4f4d<\/th>\n<th>\u914d\u7f6e<\/th>\n<th>\u4f9d\u636e<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>\u6700\u4f4e\u80fd\u7528<\/td>\n<td>4 \u6838 CPU \/ 8GB \u5185\u5b58 \/ \u65e0\u72ec\u663e<\/td>\n<td>SenseVoice 234M \u53c2\u6570\u6309 fp32 \u7ea6 0.9GB\uff0c\u52a0 VAD \u548c\u6807\u70b9\u6a21\u578b\uff0c\u6743\u91cd\u603b\u91cf 1.3GB \u5de6\u53f3\uff1b\u7eaf\u6587\u4ef6\u8f6c\u5199\u4e0d\u8ffd\u901f\u5ea6\u591f\u7528<\/td>\n<\/tr>\n<tr>\n<td>\u5efa\u8bae\u914d\u7f6e<\/td>\n<td>8 \u6838\u4ee5\u4e0a \/ 16GB \u5185\u5b58 \/ 6GB \u4ee5\u4e0a\u663e\u5b58<\/td>\n<td>\u663e\u5b58\u8fd9\u4e00\u6863\u53c2\u8003\u4e86\u7b2c\u4e09\u65b9\u6d4b\u8bc4\u7ed9\u51fa\u7684 SenseVoice + VAD + SPK \u5e38\u9a7b\u7ea6 3.8GB \u7684\u6570\u5b57\uff0c\u7559\u4e00\u500d\u4f59\u91cf<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p>\u90a3\u4e2a 3.8GB \u6765\u81ea <code>andrew.ooo<\/code> \u7684\u4e00\u7bc7\u7b2c\u4e09\u65b9\u6d4b\u8bc4\uff0c\u4e0d\u662f\u5b98\u65b9\u6570\u636e\uff0c\u6211\u5f53\u53c2\u8003\u4e0d\u5f53\u7ed3\u8bba\u3002\u540c\u4e00\u7bc7\u91cc\u63d0\u5230 faster-whisper \u8dd1 large-v3 \u5e38\u9a7b\u7ea6 2.4GB\u2014\u2014<strong>\u8fd9\u6761\u94fe\u7684\u663e\u5b58\u5360\u7528\u53cd\u800c\u66f4\u9ad8<\/strong>\uff0c\u56e0\u4e3a\u56db\u4e2a\u6a21\u578b\u540c\u65f6\u6302\u7740\u3002<\/p>\n<p>\u78c1\u76d8\u8fd9\u5757\u6211\u8dd1\u5b8c\u624d\u53bb\u770b\uff0c\u6570\u5b57\u6709\u70b9\u610f\u5916\u3002\u6743\u91cd\u7f13\u5b58 <code>~\/.cache\/modelscope<\/code> \u4e00\u5171 <strong>2.0GB<\/strong>\uff1a<code>ct-punc<\/code> 1.1GB\u3001<code>SenseVoiceSmall<\/code> 900MB\u3001<code>cam++<\/code> 28MB\u3001<code>fsmn-vad<\/code> 3.9MB\u3002<\/p>\n<p><strong>\u6807\u70b9\u6a21\u578b\u6bd4\u8bc6\u522b\u6a21\u578b\u8fd8\u5927\u3002<\/strong> \u60f3\u4e00\u4e0b\u53c8\u5408\u7406\uff1a<code>ct-punc<\/code> \u662f 290M \u53c2\u6570\u3001\u8bcd\u8868\u56db\u5341\u4e03\u4e07\u7684\u6a21\u578b\uff0cSenseVoice \u53ea\u6709 234M\u3002\u53ea\u300c\u542c\u51fa\u5b57\u300d\u4e00\u5957 900MB\uff1b\u60f3\u8ba9\u8f93\u51fa\u5e26\u6807\u70b9\u3001\u80fd\u5f53\u7eaa\u8981\u8bfb\uff0c\u78c1\u76d8\u548c\u5185\u5b58\u7acb\u523b\u7ffb\u500d\u3002\u8fd9\u5c31\u662f\u300c\u4e2d\u6587\u65ad\u53e5\u9760\u6807\u70b9\u6a21\u578b\u300d\u7684\u7269\u7406\u6210\u672c\u3002<\/p>\n<p>\u6a21\u578b\u6258\u7ba1\u5728 ModelScope\uff1b\u5b98\u65b9\u6392\u969c\u6587\u6863\u8bf4\u4e2d\u56fd\u5927\u9646\u8d70 ModelScope \u66f4\u5feb\uff0c\u5883\u5916\u53ef\u4ee5\u4f20 <code>--hub hf<\/code>\u3002\u5883\u5916\u8bbf\u95ee ModelScope \u7684 CDN \u6162\u662f\u957f\u671f\u95ee\u9898\u3002<\/p>\n<h2 id=\"json\">\u6211\u4ece\u547d\u4ee4\u884c\u4e00\u8def\u8d70\u5230 JSON<\/h2>\n<p>\u7aef\u5230\u7aef\u6700\u7701\u4e8b\u7684\u8def\u5f84\u662f CLI\u3002\u6211\u7528\u5b98\u65b9\u7ed9\u7684\u516c\u5f00\u4e2d\u6587\u6837\u4f8b\u97f3\u9891\uff08<code>asr_example_zh.wav<\/code>\uff0c5.55 \u79d2\u300116kHz \u5355\u58f0\u9053\uff09\u8dd1\u7684\uff0c\u547d\u4ee4\u91cc\u628a\u4e09\u4e2a\u5f00\u5173\u90fd\u6253\u5f00\uff1a<\/p>\n<pre><code class=\"language-bash\">funasr asr_example_zh.wav --spk --timestamps -f json -v\n<\/code><\/pre>\n<p><code>-v<\/code> \u5fc5\u987b\u52a0\u3002\u4e0d\u52a0\u4f60\u9762\u5bf9\u7684\u662f\u4e00\u6bb5\u6ca1\u6709\u8f93\u51fa\u7684\u6c89\u9ed8\uff0c\u4e0d\u77e5\u9053\u5b83\u5728\u4e0b\u8f7d\u6743\u91cd\u8fd8\u662f\u5728\u63a8\u7406\u3002\u52a0\u4e0a\u4e4b\u540e\uff0cstderr \u4f1a\u7ed9\u4f60 <code>Model loaded in 88.1s<\/code> \u548c\u6bcf\u4e00\u7ea7\u7684 rtf\u3002<\/p>\n<p>\u6211\u8fd9\u53f0 M2 \u4e0a\u7684\u771f\u5b9e\u8017\u65f6\uff1a<\/p>\n<table>\n<thead>\n<tr>\n<th>\u573a\u666f<\/th>\n<th>\u5899\u949f\u65f6\u95f4<\/th>\n<th>\u5176\u4e2d\u6a21\u578b\u52a0\u8f7d<\/th>\n<th>\u7eaf\u63a8\u7406<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>\u9996\u6b21\uff08\u8981\u4e0b\u8f7d\u6743\u91cd\uff09<\/td>\n<td>95.12s<\/td>\n<td>88.1s<\/td>\n<td>1.243s<\/td>\n<\/tr>\n<tr>\n<td>\u4e8c\u6b21\uff08\u6743\u91cd\u5df2\u7f13\u5b58\uff09<\/td>\n<td>13.02s<\/td>\n<td>\u7ea6 11.5s<\/td>\n<td>1.528s<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p>5.55 \u79d2\u97f3\u9891\u30011.2 \u5230 1.5 \u79d2\u51fa\u7ed3\u679c\uff0crtf 0.21 \u4e0a\u4e0b\uff1bstderr \u91cc ASR \u90a3\u4e00\u7ea7\u7684 <code>rtf<\/code> \u662f 0.016\uff0c\u5dee\u4e00\u4e2a\u91cf\u7ea7\u2014\u2014\u6574\u6761\u94fe\u8fd8\u8981\u8dd1 VAD \u548c\u8bf4\u8bdd\u4eba\u805a\u7c7b\uff0c\u8fd9\u4e24\u6b65\u8017\u65f6\u56fa\u5b9a\uff0c\u97f3\u9891\u8d8a\u77ed\u644a\u5f97\u8d8a\u96be\u770b\u3002<\/p>\n<p>\u8bc6\u522b\u7ed3\u679c\u548c\u5b98\u65b9 README \u91cc\u5199\u7684\u90a3\u53e5\u6837\u4f8b\u6587\u672c\u4e00\u81f4\uff1a<code>\u6b22\u8fce\u5927\u5bb6\u6765\u4f53\u9a8c\u8fbe\u6469\u9662\u63a8\u51fa\u7684\u8bed\u97f3\u8bc6\u522b\u6a21\u578b<\/code>\u3002\u5341\u4e5d\u4e2a\u5b57\uff0c\u4e00\u4e2a\u6ca1\u9519\u3002<\/p>\n<p>\u8f93\u51fa\u7684 JSON \u662f\u8fd9\u6837\uff08\u6211\u5220\u6389\u4e86\u4e2d\u95f4\u5341\u4e03\u4e2a\u65f6\u95f4\u6233\uff09\uff1a<\/p>\n<pre><code class=\"language-json\">{\n  &quot;text&quot;: &quot;\u6b22\u8fce\u5927\u5bb6\u6765\u4f53\u9a8c\u8fbe\u6469\u9662\u63a8\u51fa\u7684\u8bed\u97f3\u8bc6\u522b\u6a21\u578b&quot;,\n  &quot;segments&quot;: [\n    {\n      &quot;start&quot;: 610,\n      &quot;end&quot;: 5530,\n      &quot;text&quot;: &quot;\u6b22\u8fce\u5927\u5bb6\u6765\u4f53\u9a8c\u8fbe\u6469\u9662\u63a8\u51fa\u7684\u8bed\u97f3\u8bc6\u522b\u6a21\u578b&quot;,\n      &quot;timestamp&quot;: [[940, 1000], [1120, 1180], &quot;\u2026\u2026\u5171 19 \u7ec4\u2026\u2026&quot;],\n      &quot;speaker&quot;: 0\n    }\n  ],\n  &quot;file&quot;: &quot;asr_example_zh.wav&quot;,\n  &quot;model&quot;: &quot;sensevoice&quot;,\n  &quot;language&quot;: &quot;auto&quot;,\n  &quot;audio_duration_s&quot;: 5.547,\n  &quot;processing_s&quot;: 1.243\n}\n<\/code><\/pre>\n<p>\u4e09\u4e2a\u7ec6\u8282\u3002<code>start<\/code>\u3001<code>end<\/code>\u3001<code>timestamp<\/code> \u91cc\u7684\u6570\uff0c\u5355\u4f4d\u662f<strong>\u6beb\u79d2<\/strong>\u3002<code>speaker<\/code> \u662f\u5f55\u97f3\u5185\u90e8\u7f16\u53f7\uff0c\u4ece 0 \u5f00\u59cb\uff0c\u6362\u6587\u4ef6\u91cd\u65b0\u6392\uff0c\u4e0d\u4ee3\u8868\u67d0\u4e2a\u5177\u4f53\u7684\u4eba\u2014\u2014\u5b98\u65b9\u6587\u6863\u5199\u5f97\u5f88\u6b7b\uff0c\u300c\u5f55\u97f3\u5185\u533f\u540d\u6807\u7b7e\uff0c\u4e0d\u662f\u5df2\u6ce8\u518c\u8bf4\u8bdd\u4eba\u8bc6\u522b\u300d\uff0c\u60f3\u505a\u300c\u8fd9\u662f\u5f20\u7ecf\u7406\u300d\u5f97\u81ea\u5df1\u52a0\u4e00\u5c42\u58f0\u7eb9\u6ce8\u518c\uff0c<code>cam++<\/code> \u53ea\u7ed9\u5d4c\u5165\u5411\u91cf\u3002<code>processing_s<\/code> \u662f\u7eaf\u63a8\u7406\u8017\u65f6\uff0c\u62ff\u5b83\u7b97\u6210\u672c\u6bd4\u5899\u949f\u51c6\u3002<\/p>\n<p>\u7136\u540e\u662f\u8fd9\u4e00\u8d9f\u91cc\u6211\u6700\u60f3\u8ba9\u4eba\u63d0\u524d\u77e5\u9053\u7684\u4e00\u4ef6\u4e8b\uff1a<strong><code>--spk<\/code> \u8fd9\u4e2a\u5f00\u5173\uff0c\u547d\u4ee4\u884c\u5e76\u6ca1\u6709\u987a\u624b\u628a\u6807\u70b9\u6a21\u578b\u5e26\u4e0a\u3002<\/strong> \u6211\u7684 stderr \u91cc\u8eba\u7740\u8fd9\u4e48\u4e00\u884c\uff1a<\/p>\n<pre><code class=\"language-text\">[WARNING] punc_model is missing, falling back to vad_segment mode for speaker diarization.\n<\/code><\/pre>\n<p>\u610f\u601d\u662f\u8bf4\u8bdd\u4eba\u5206\u79bb\u9000\u56de\u4e86\u300c\u6309 VAD \u6bb5\u5224\u5b9a\u300d\u2014\u2014\u53ea\u77e5\u9053\u8fd9\u51e0\u6bb5\u4e4b\u95f4\u6709\u505c\u987f\uff0c\u4e0d\u77e5\u9053\u6807\u70b9\u5728\u54ea\uff0c\u4e8e\u662f\u8bf4\u8bdd\u4eba\u5207\u6362\u70b9\u53ea\u80fd\u843d\u5728\u505c\u987f\u4e0a\u3002\u4e24\u4e2a\u4eba\u62a2\u8bdd\u3001\u6216\u8005\u4e00\u4e2a\u4eba\u8bf4\u5b8c\u53e6\u4e00\u4e2a\u9a6c\u4e0a\u63a5\uff0c\u8fd9\u4e00\u6bb5\u5c31\u4f1a\u88ab\u5224\u6210\u540c\u4e00\u4e2a\u4eba\u3002\u60f3\u8981\u5e72\u51c0\u7684\u5207\u6362\uff0c\u5f97\u5728 Python \u91cc\u663e\u5f0f\u6302\u4e0a <code>punc_model=\"ct-punc\"<\/code>\uff0cCLI \u6ca1\u7ed9\u4f60\u8fd9\u4e2a\u53e3\u5b50\u3002\u8fd9\u6761\u662f\u6211\u81ea\u5df1\u649e\u51fa\u6765\u7684\u3002<\/p>\n<p>\u987a\u5e26\u4e24\u4e2a\u5c0f\u63d0\u793a\u3002\u4e00\u662f\u542f\u52a8\u65f6\u4f1a\u544a\u8bc9\u4f60 <code>ffmpeg is not installed. torchaudio is used to load audio<\/code>\u2014\u2014\u6ca1\u88c5\u4e0d\u4f1a\u5931\u8d25\uff0c\u53ea\u662f\u6362 torchaudio \u8bfb\u97f3\u9891\uff0cmp3\u3001m4a \u517c\u5bb9\u6027\u5dee\u4e00\u70b9\u3002\u4e8c\u662f SRT \u8f93\u51fa\uff0c<code>funasr asr_example_zh.wav -f srt -o .\/subs<\/code> \u7ed9\u6211\u7684\u662f\u5355\u6761\u5b57\u5e55\uff1a<\/p>\n<pre><code class=\"language-text\">1\n00:00:00,940 --&gt; 00:00:04,900\n\u6b22\u8fce\u5927\u5bb6\u6765\u4f53\u9a8c\u8fbe\u6469\u9662\u63a8\u51fa\u7684\u8bed\u97f3\u8bc6\u522b\u6a21\u578b\u3002\n<\/code><\/pre>\n<p>\u6ce8\u610f\u5b83\u81ea\u52a8\u8865\u4e86\u4e2a\u53e5\u53f7\uff0c\u4e5f\u628a\u8d77\u59cb\u65f6\u95f4\u5bf9\u5230\u4e86\u7b2c\u4e00\u4e2a\u5b57\u7684 940 \u6beb\u79d2\uff0c\u800c\u4e0d\u662f\u4ece 0 \u5f00\u59cb\u2014\u2014\u8fd9\u4e2a\u7ec6\u8282\u505a\u5f97\u633a\u5bf9\uff0c\u5f88\u591a\u5de5\u5177\u4f1a\u4ece 0 \u8d77\u3002<\/p>\n<p>\u60f3\u8981\u66f4\u50cf\u4f1a\u8bae\u7eaa\u8981\u7684\u6392\u7248\u3001\u5e76\u4e14\u628a SenseVoice \u7684 <code>&lt;|HAPPY|&gt;<\/code> \u8fd9\u7c7b\u6807\u7b7e\u5265\u6389\uff0c\u53ea\u80fd\u56de\u5230 Python\uff1a<\/p>\n<pre><code class=\"language-python\">from funasr import AutoModel\nfrom funasr.utils.postprocess_utils import rich_transcription_postprocess\n\nmodel = AutoModel(\n    model=&quot;iic\/SenseVoiceSmall&quot;,\n    vad_model=&quot;fsmn-vad&quot;,\n    punc_model=&quot;ct-punc&quot;,      # \u4e0d\u52a0\u8fd9\u4e2a\uff0c\u8bf4\u8bdd\u4eba\u5207\u6362\u70b9\u4f1a\u5f88\u7c97\n    spk_model=&quot;cam++&quot;,\n    device=&quot;cpu&quot;,\n)\nresult = model.generate(input=&quot;meeting.wav&quot;, batch_size_s=300)\n\nfor seg in result[0][&quot;sentence_info&quot;]:\n    print(f&quot;[{seg['start']\/1000:.1f}s] \u8bf4\u8bdd\u4eba{seg['spk']}: &quot;\n          f&quot;{rich_transcription_postprocess(seg['sentence'])}&quot;)\n<\/code><\/pre>\n<p>\u8fd9\u91cc <code>sentence_info<\/code> \u7684\u5b57\u6bb5\u662f <code>start<\/code>\u3001<code>end<\/code>\u3001<code>spk<\/code>\u3001<code>sentence<\/code>\uff0c\u548c CLI \u7684 <code>segments<\/code> \u4e0d\u662f\u4e00\u5957\u547d\u540d\u2014\u2014\u540c\u4e00\u4e2a\u9879\u76ee\uff0c\u4e24\u6761\u5165\u53e3\u4e24\u5957\u7ed3\u6784\u3002<code>batch_size_s=300<\/code> \u662f\u300c\u4e00\u6b21\u62fc\u4e09\u767e\u79d2\u97f3\u9891\u9001\u8fdb\u6a21\u578b\u300d\uff0c\u4e0d\u662f\u6279\u6b21\u6570\uff0c\u8c03\u5927\u6709\u541e\u5410\u6536\u76ca\u3001\u4ee3\u4ef7\u662f\u5185\u5b58\u7ebf\u6027\u4e0a\u5347\uff1b\u8fd9\u4e2a\u53c2\u6570\u6211\u6ca1\u5728\u957f\u97f3\u9891\u4e0a\u538b\u8fc7\uff0c\u65cb\u5230\u591a\u5c11\u5f97\u770b\u4f60\u7684\u673a\u5668\u3002<\/p>\n<p>\u6d41\u5f0f\u90a3\u6761\u8def\u662f\u53e6\u4e00\u5957\u4e1c\u897f\uff0c\u522b\u6df7\u7740\u7528\u3002\u5b83\u9700\u8981\u4e13\u95e8\u7684\u6d41\u5f0f\u6a21\u578b\uff0c\u4ee3\u7801\u957f\u8fd9\u6837\uff1a<\/p>\n<pre><code class=\"language-python\">import soundfile as sf\nfrom funasr import AutoModel\n\nmodel = AutoModel(model=&quot;paraformer-zh-streaming&quot;, device=&quot;cpu&quot;)\naudio, sr = sf.read(&quot;speech.wav&quot;, dtype=&quot;float32&quot;)   # \u5fc5\u987b 16kHz \u5355\u58f0\u9053\n\nchunk_size = [0, 10, 5]                              # 600ms \u4e00\u5757\nstride = chunk_size[1] * 960\ncache = {}\nn = (len(audio) - 1) \/\/ stride + 1\nfor i in range(n):\n    chunk = audio[i*stride:(i+1)*stride]\n    res = model.generate(input=chunk, cache=cache, is_final=(i == n-1),\n                         chunk_size=chunk_size,\n                         encoder_chunk_look_back=4, decoder_chunk_look_back=1)\n    if res[0][&quot;text&quot;]:\n        print(res[0][&quot;text&quot;], end=&quot;&quot;, flush=True)\n<\/code><\/pre>\n<p>\u4e24\u4e2a <code>look_back<\/code> \u662f\u7cbe\u9ad3\uff1a<code>encoder_chunk_look_back=4<\/code> \u662f\u300c\u7f16\u7801\u5668\u5f80\u524d\u770b\u56db\u5757\u300d\uff0c<code>decoder_chunk_look_back=1<\/code> \u662f\u300c\u89e3\u7801\u5668\u5f80\u524d\u770b\u4e00\u5757\u300d\u3002\u5f80\u524d\u770b\u5f97\u591a\uff0c\u8fb9\u754c\u5904\u7684\u5b57\u5c31\u51c6\uff0c\u4ee3\u4ef7\u662f\u5ef6\u8fdf\u53d8\u9ad8\u3002<code>[0, 10, 5]<\/code> \u662f\u516d\u767e\u6beb\u79d2\u4e00\u5757\uff0c\u6709\u4eba\u5acc\u5b83\u5bf9\u76f4\u64ad\u592a\u4fdd\u5b88\uff0c\u60f3\u66f4\u4f4e\u5ef6\u8fdf\u5c31\u628a\u5757\u5207\u5c0f\uff0c\u51c6\u786e\u7387\u4f1a\u6389\u3002\u8fd9\u65cb\u94ae\u6ca1\u6709\u6807\u51c6\u7b54\u6848\u3002<\/p>\n<h2 id=\"issue\">issue \u533a\u91cc\u7ffb\u5230\u7684\u516d\u4ef6\u4e8b<\/h2>\n<p>\u516c\u5f00\u8ba8\u8bba\u8fd9\u5757\uff0c\u4e2d\u6587\u5708\u8c08 FunASR \u7684\u6587\u7ae0\u4e0d\u5c11\uff0c\u591a\u6570\u662f\u300c\u901f\u5ea6\u5feb\u3001\u4e2d\u6587\u51c6\u300d\u7684\u590d\u8ff0\u3002\u6211\u6311\u51e0\u4e2a\u6709\u4fe1\u606f\u91cf\u7684\u8bf4\uff0c\u90fd\u80fd\u70b9\u8fdb\u53bb\u6838\u3002<\/p>\n<p><strong>\u70ed\u8bcd\u4f1a\u8ba9\u53e5\u5b50\u53d8\u77ed\u3002<\/strong> \u8fd9\u662f issue \u533a\u8ba8\u8bba\u6700\u591a\u7684\u4e00\u4e2a\uff0c\u7f16\u53f7 #3302\uff0c34 \u6761\u8bc4\u8bba\uff0c\u4ece 7 \u6708\u5f00\u5230\u4eca\u5929\u6ca1\u5173\u3002\u63d0\u95ee\u7684\u4eba\u63cf\u8ff0\u5f97\u5f88\u5177\u4f53\uff1a\u5b9e\u65f6\u670d\u52a1\u91cc\uff0c\u9884\u89c8\uff08partial\uff09\u80fd\u770b\u5230\u7684\u6587\u5b57\uff0c\u5230\u6700\u540e\u786e\u8ba4\uff08final\uff09\u90a3\u7248\u4f1a\u5c11\u6389\u4e00\u5927\u622a\uff1b\u70ed\u8bcd\u52a0\u5f97\u8d8a\u591a\u4e22\u5f97\u8d8a\u5389\u5bb3\uff0c\u8fd8\u51fa\u73b0\u8fc7\u300c\u9759\u97f3\u72b6\u6001\u4e0b\u5192\u51fa\u70ed\u8bcd\u91cc\u7684\u5b57\u300d\u3002\u6211\u7684\u7406\u89e3\u662f\u70ed\u8bcd\u66ff\u6362\u53d1\u751f\u5728\u5206\u6bb5\u7ed3\u679c\u4e0a\uff0c\u66ff\u6362\u4e00\u65e6\u6539\u53d8\u53e5\u5b50\u7ed3\u6784\uff0c\u8fb9\u754c\u5c31\u8ddf\u7740\u91cd\u7b97\uff0c\u91cd\u7b97\u5c31\u4f1a\u5403\u6389\u5185\u5bb9\u3002\u8fd9\u6761\u81f3\u4eca\u6ca1\u6709\u5b98\u65b9\u7ed3\u8bba\uff0c<strong>\u6211\u7684\u5efa\u8bae\u662f\u79bb\u7ebf\u7528\u70ed\u8bcd\uff0c\u5b9e\u65f6\u522b\u7528<\/strong>\u3002<\/p>\n<p><strong>GGUF \u7684 Vulkan \u540e\u7aef\u5728 AMD \u5361\u4e0a\u5d29\u3002<\/strong> #3479\uff0c22 \u6761\u8bc4\u8bba\uff0cWindows 11 + RX 7600M XT\uff0c<code>--backend vulkan<\/code> \u5728\u8bbe\u5907\u68c0\u6d4b\u4e4b\u540e\u7acb\u523b <code>0xc0000005<\/code> \u8bbf\u95ee\u51b2\u7a81\uff0c\u62ff <code>sensevoice-small-f16.gguf<\/code> \u80fd\u590d\u73b0\u3002\u5b98\u65b9\u5728 issue \u91cc\u8981\u6c42\u63d0\u4f9b <code>nvidia-smi<\/code>\uff0c\u53ef\u8fd9\u662f AMD \u5361\uff0c\u672c\u6765\u5c31\u6ca1\u6709\u2014\u2014\u6a21\u677f\u5316\u56de\u590d\uff0c\u7b97\u662f\u56e2\u961f\u5de5\u7a0b\u4e60\u60ef\u4e0a\u7684\u4e00\u4e2a\u7c97\u7cd9\u9762\u3002<\/p>\n<p><strong>\u5b9e\u65f6\u670d\u52a1\u8fbe\u4e0d\u5230\u5b9e\u65f6\u3002<\/strong> #3038\uff0c16 \u6761\u8bc4\u8bba\uff0c\u73af\u5883\u662f L20 \u5355\u5361\uff0c\u6d4b\u8bd5\u65b9\u6cd5\u662f\u6d4f\u89c8\u5668\u5f00\u591a\u4e2a demo \u540c\u65f6\u653e B \u7ad9\u957f\u97f3\u9891\uff0c\u7ed3\u8bba\u662f\u300c\u64ad\u5f97\u8d8a\u4e45\u5ef6\u8fdf\u8d8a\u5927\u300d\u3002\u8fd9\u4e2a\u73b0\u8c61\u5f88\u5178\u578b\uff1a\u4e0d\u662f\u5355\u5757\u63a8\u7406\u6162\uff0c\u662f\u7f13\u5b58\u548c\u961f\u5217\u5728\u7d2f\u52a0\u3002\u60f3\u4e0a\u5b9e\u65f6\u5b57\u5e55\u7684\u4eba\uff0c\u8fd9\u6761\u5fc5\u8bfb\u3002<\/p>\n<p><strong>CLI \u548c\u4ee3\u7801\u8dd1\u51fa\u6765\u7684\u7ed3\u679c\u4e0d\u4e00\u6837\u3002<\/strong> #3773\uff0c10 \u6708 14 \u65e5\u63d0\u7684\u3002\u63d0\u95ee\u7684\u4eba\u53d1\u73b0\u540c\u4e00\u4e2a\u97f3\u9891\u3001\u540c\u4e00\u7248 1.4.16\uff0c\u7528 <code>funasr<\/code> \u547d\u4ee4\u8dd1\u51fa\u6765\u7684\u65ad\u53e5\u6bd4 Python \u91cc\u8c03 <code>AutoModel<\/code> \u66f4\u5408\u7406\u3002\u6211\u987a\u7740\u770b\u4e86\u6e90\u7801\uff1a<code>cli.py<\/code> \u751f\u6210\u5b57\u5e55\u65f6\u4f1a\u8c03 <code>jieba<\/code> \u5206\u8bcd\uff0c\u6309\u8bcd\u8fb9\u754c\u51b3\u5b9a\u5728\u54ea\u91cc\u65ad\u5f00\uff1b\u6ca1\u6709 <code>jieba<\/code> \u5c31\u9000\u5316\u6210\u53ea\u6309\u6807\u70b9\u65ad\uff0c\u4ee3\u7801\u91cc\u4f1a\u6253 warning \u8bf4\u300c\u53ef\u80fd\u628a\u8bcd\u5207\u5230\u4e24\u884c\u53bb\u300d\uff0c10 \u6708 27 \u65e5\u6709\u4e2a\u63d0\u4ea4\uff08#3726\uff09\u4e13\u95e8\u7ed9\u8fd9\u4e2a\u964d\u7ea7\u52a0\u63d0\u793a\u3002\u4e0d\u8fc7 <code>jieba<\/code> \u662f\u9ed8\u8ba4\u4f9d\u8d56\uff0c<code>pip install funasr<\/code> \u88c5\u5b8c\u5c31\u6709\uff0c\u6240\u4ee5\u5927\u591a\u6570\u4eba\u770b\u4e0d\u5230\u8fd9\u6761 warning\u2014\u2014\u771f\u6b63\u4f1a\u9000\u5316\u7684\uff0c\u662f\u624b\u5de5\u88c1\u4f9d\u8d56\u7684\u7cbe\u7b80\u955c\u50cf\u3002\u7eaf\u6587\u672c\u8f93\u51fa\u4e24\u6761\u8def\u6709\u6ca1\u6709\u533a\u522b\uff0c\u6211\u6ca1\u9010\u53e5\u6bd4\u8fc7\uff0c\u4e0d\u4e0b\u7ed3\u8bba\u3002<\/p>\n<p><strong>\u6743\u91cd\u8bb8\u53ef\u548c\u4ee3\u7801\u8bb8\u53ef\u662f\u4e24\u4ef6\u4e8b\u3002<\/strong> \u4ed3\u5e93\u662f MIT\uff0c\u4f46\u6a21\u578b\u6743\u91cd\u53e6\u6709\u8bb8\u53ef\u3002#3458 \u8fd9\u6761\u95ee\u7684\u5c31\u662f\u300c\u56db\u4e2a\u56fa\u5b9a revision \u7684\u6a21\u578b\u5728\u81ea\u6258\u7ba1\u548c\u6743\u91cd\u5206\u53d1\u573a\u666f\u4e0b\u5230\u5e95\u9002\u7528\u4ec0\u4e48\u6761\u6b3e\u300d\uff0c\u56de\u7b54\u5e76\u4e0d\u5e72\u8106\u3002<strong>\u5982\u679c\u4f60\u7684\u573a\u666f\u662f\u628a\u6743\u91cd\u6253\u8fdb\u79c1\u6709\u90e8\u7f72\u5305\u91cc\u53d1\u7ed9\u5ba2\u6237\uff0c\u8fd9\u4e00\u6761\u5fc5\u987b\u81ea\u5df1\u627e\u6cd5\u52a1\u786e\u8ba4<\/strong>\uff0c\u522b\u770b\u5230 MIT \u5c31\u4ee5\u4e3a\u6574\u6761\u94fe\u90fd\u80fd\u968f\u4fbf\u5206\u53d1\u3002<\/p>\n<p>\u8fd8\u6709\u4e00\u6761\u5bf9\u505a\u65f6\u95f4\u6233\u7684\u4eba\u5f88\u81f4\u547d\uff1a#3496\uff0cFun-ASR-Nano \u7684 Hugging Face checkpoint \u7f3a\u4e86 86 \u4e2a CTC \u5f20\u91cf\uff0c\u65f6\u95f4\u6233\u548c\u8bf4\u8bdd\u4eba\u5206\u79bb\u90a3\u6761\u8def\u8d70\u4e0d\u901a\u3002<code>-hf<\/code> \u90a3\u4e2a\u6743\u91cd\u548c FunASR \u7248<strong>\u4e0d\u662f\u7b49\u4ef7\u7684\u5305<\/strong>\uff0c\u9009\u9519\u4e00\u4e2a\u529f\u80fd\u5c31\u5c11\u4e00\u5757\u3002<\/p>\n<p>\u5916\u90e8\u89c6\u89d2\u4e5f\u627e\u4e86\u51e0\u4e2a\u3002<code>funasr.com<\/code> \u5b98\u7f51\u6709\u4e00\u7bc7\u81ea\u5bb6\u505a\u7684\u5bf9\u6bd4\uff0c184 \u6bb5\u4e2d\u6587\u97f3\u9891\u3001\u540c\u673a CPU\u3001\u5b57\u7b26\u7ea7 CER\uff1aSenseVoice 8.0%\u3001Paraformer 9.9%\u3001Fun-ASR-Nano 8.3%\uff0cwhisper \u7cfb\uff08small\/base\/large-v3-turbo\uff09\u5728 22% \u5230 31% \u4e4b\u95f4\u3002\u5b98\u65b9\u81ea\u6d4b\uff0c\u53e3\u5f84\u5bf9\u5b83\u6709\u5229\uff0c\u6211\u5f53\u300c\u65b9\u5411\u53ef\u4fe1\u3001\u7edd\u5bf9\u503c\u6253\u516b\u6298\u300d\u3002<code>hysenlabs.com<\/code> \u6709\u4e00\u7bc7\u504f\u5de5\u7a0b\u7684\u89e3\u8bfb\uff0c\u63d0\u5230\u4e24\u4e2a\u6211\u6ca1\u6ce8\u610f\u7684 extras\uff1a<code>funasr[silero]<\/code> \u6362 VAD \u540e\u7aef\uff0c<code>kaldi-native-fbank<\/code> \u662f\u7ed9\u6ca1\u6709 torchaudio \u8f6e\u5b50\u7684\u6607\u817e\u3001aarch64 \u670d\u52a1\u5668\u515c\u5e95\u2014\u2014\u8fd9\u4e24\u4e2a extra \u7684\u5b58\u5728\u8bf4\u660e\uff0c\u5b83\u7684\u9ed8\u8ba4\u5047\u8bbe\u662f\u300c\u4e00\u53f0\u6b63\u5e38\u7684 x86 + torchaudio \u673a\u5668\u300d\u3002<\/p>\n<h2 id=\"_3\">\u8c01\u8be5\u642c\uff0c\u8c01\u522b\u52a8<\/h2>\n<p><strong>\u503c\u5f97\u642c\u7684<\/strong>\uff1a\u5f55\u97f3\u8981\u7559\u5728\u81ea\u5df1\u673a\u5668\u91cc\u7684\uff08\u4f1a\u8bae\u3001\u8bbf\u8c08\u3001\u5ba2\u670d\u3001\u533b\u7597\uff09\uff1b\u4e2d\u6587\u6216\u4e2d\u82f1\u6df7\u8bf4\u4e3a\u4e3b\u7684\uff1b\u9700\u8981\u300c\u8c01\u5728\u4ec0\u4e48\u65f6\u5019\u8bf4\u4e86\u4ec0\u4e48\u300d\u8fd9\u79cd\u5e26\u8bf4\u8bdd\u4eba\u548c\u65f6\u95f4\u6233\u7684\u7ed3\u6784\u5316\u8f93\u51fa\uff1b\u624b\u4e0a\u5df2\u6709\u4e00\u5806\u5de5\u5177\u3001\u53ea\u7f3a\u4e00\u4e2a\u672c\u5730 <code>base_url<\/code> \u7684\u2014\u2014\u8fd9\u4e00\u7c7b\u6536\u76ca\u6700\u5927\uff0c\u63a5\u4e0a\u5c31\u80fd\u7528\u3002<\/p>\n<p><strong>\u522b\u52a8\u7684<\/strong>\uff1a\u8981\u5c0f\u8bed\u79cd\u548c\u7ffb\u8bd1\u7684\uff0c\u53bb\u7528 whisper \u7cfb\uff1b\u8981\u8bcd\u7ea7\u7cbe\u786e\u5bf9\u9f50\u505a\u9010\u5b57\u5b57\u5e55\u7684\uff0c\u8fd9\u4e0d\u662f\u5b83\u5f3a\u9879\uff1b\u663e\u5b58\u53ea\u6709 4GB \u8fd8\u60f3\u6302 VAD + \u8bf4\u8bdd\u4eba\u7684\uff0c\u4f1a\u96be\u53d7\uff1b\u4ee5\u53ca\u6700\u91cd\u8981\u7684\u2014\u2014<strong>\u8981\u300c\u88c5\u5b8c\u5c31\u80fd\u7528\u3001\u51fa\u9519\u6709\u4eba\u7ba1\u300d\u7684\uff0c\u522b\u78b0<\/strong>\u3002\u8fd9\u662f\u4e2a\u5de5\u5177\u5305\u4e0d\u662f\u4ea7\u54c1\uff0c\u5b83\u7684\u6587\u6863\u4f1a\u660e\u786e\u5199\u300c\u8dd1\u901a\u4e0d\u4ee3\u8868\u7ed3\u679c\u53ef\u4fe1\u300d\uff0c\u5269\u4e0b\u5168\u662f\u4f60\u81ea\u5df1\u7684\u5224\u65ad\u3002<\/p>\n<p><strong>\u66ff\u4ee3\u65b9\u6848<\/strong>\uff0c\u6211\u62ff\u4e24\u4e2a\u6bd4\uff1a<\/p>\n<p><strong>faster-whisper<\/strong>\uff08<code>SYSTRAN\/faster-whisper<\/code>\uff0c26,235 star\uff09\u3002\u5b83\u662f whisper \u7684 CTranslate2 \u91cd\u5199\u7248\uff0c\u7ef4\u5ea6\u5dee\u522b\u5728<strong>\u591a\u8bed\u8a00\u4e0e\u751f\u6001<\/strong>\uff1a99 \u79cd\u8bed\u8a00\u3001\u80fd\u7ffb\u8bd1\u3001\u793e\u533a\u5927\u3001\u8e29\u5751\u8d44\u6599\u591a\uff0c\u4ee3\u4ef7\u662f\u4e2d\u6587 CER \u9ad8\u51fa\u4e24\u4e09\u500d\u3001\u663e\u5b58\u5e38\u9a7b\u53cd\u800c\u66f4\u4f4e\u4f46\u901f\u5ea6\u6162\u5f97\u591a\u3002\u9009\u5b83\u7684\u7406\u7531\u53ea\u6709\u4e00\u4e2a\uff1a\u4f60\u7684\u97f3\u9891\u4e0d\u662f\u4e2d\u6587\u4e3a\u4e3b\u3002<\/p>\n<p><strong>whisper.cpp<\/strong>\uff08<code>ggml-org\/whisper.cpp<\/code>\uff0c54,713 star\uff09\u3002\u7ef4\u5ea6\u5dee\u522b\u5728<strong>\u90e8\u7f72\u5f62\u6001<\/strong>\uff1a\u4e00\u4e2a\u4e8c\u8fdb\u5236\u3001\u96f6 Python\u3001Mac \u4e0a\u8fd8\u80fd\u5403 Metal\u3002FunASR \u4e5f\u505a\u4e86\u5bf9\u5e94\u7684 GGUF \u8fd0\u884c\u65f6\uff0c\u4f46\u5b98\u65b9 release \u91cc\u6211\u53ea\u770b\u5230 Linux x64 \u7684 Vulkan \u5305\u548c Windows \u7684 Vulkan\/CUDA \u5305\uff0c<strong>\u6ca1\u6709 macOS \u7684<\/strong>\u2014\u2014\u60f3\u5728 Mac \u4e0a\u8dd1\u5355\u6587\u4ef6\u7248\uff0c\u8fd9\u6761\u8def\u6682\u65f6\u8d70\u4e0d\u901a\u3002\u8fd9\u4e5f\u662f\u6211\u7559\u5728 Python \u7248\u7684\u539f\u56e0\u3002<\/p>\n<p>\u4e00\u53e5\u8bdd\uff1a<strong>FunASR \u6cbb\u7684\u662f\u300c\u4e2d\u6587\u542c\u4e0d\u51c6\u300d\uff0c\u4e0d\u6cbb\u300c\u88c5\u8d77\u6765\u9ebb\u70e6\u300d\u3002<\/strong><\/p>\n<p>\u603b\u800c\u8a00\u4e4b\uff0c\u5b83\u503c\u5f97\u6211\u6362\u6389\u539f\u6765\u7684 whisper.cpp \u6d41\u6c34\u7ebf\uff0c\u4ee3\u4ef7\u662f\u8981\u63a5\u53d7\u5b83\u7684\u5f62\u72b6\uff1a\u56db\u9897\u6a21\u578b\u3001\u4e00\u5806\u4f9d\u8d56\u3001\u4e00\u5806\u81ea\u5df1\u8c03\u7684\u65cb\u94ae\u3002\u987a\u624b\u770b\u4e86\u4e0b\u4ed3\u5e93\uff0c\u73b0\u5728\u662f 20,845 \u4e2a star\u30012,095 \u4e2a fork\u300158 \u4e2a\u5f85\u5904\u7406 issue\uff0c\u6700\u8fd1\u56db\u5468 43 \u6b21\u63d0\u4ea4\uff0c\u6700\u65b0\u53d1\u7248 <code>v1.4.16<\/code> \u662f 9 \u6708 18 \u65e5\uff0crelease note \u91cc\u5199\u7684\u662f\u4fee\u6389\u300cVAD\u3001\u6807\u70b9\u3001\u8bf4\u8bdd\u4eba\u6a21\u578b\u7684\u8bbe\u5907\u653e\u7f6e\u8de8\u8c03\u7528\u4e0d\u4fdd\u6301\u300d\u8fd9\u7c7b\u95ee\u9898\u2014\u2014\u90fd\u662f\u5f88\u5177\u4f53\u7684\u5de5\u7a0b\u503a\u3002<\/p>\n<p>\u6700\u540e\u8bf4\u53e5\u5b9e\u5728\u7684\uff0c\u8fd9\u4e00\u8d9f\u6211\u53ea\u9a8c\u5230\u300c5.55 \u79d2\u7684\u6837\u4f8b\u80fd\u5e72\u51c0\u5730\u8dd1\u51fa\u5e26\u65f6\u95f4\u6233\u548c\u8bf4\u8bdd\u4eba\u7684 JSON\u300d\u3002\u6211\u624b\u4e0a\u90a3\u51e0\u4e2a\u5c0f\u65f6\u7684\u771f\u5b9e\u4f1a\u8bae\u5f55\u97f3\u4e0d\u65b9\u4fbf\u5f80\u4efb\u4f55\u73af\u5883\u91cc\u4f20\uff0c\u6240\u4ee5\u957f\u97f3\u9891\u7684\u7a33\u5b9a\u6027\u548c\u62a2\u8bdd\u65f6\u7684\u5206\u79bb\u6548\u679c\uff0c\u6211\u4e00\u4e2a\u90fd\u6ca1\u6d4b\uff1bGPU \u541e\u5410\u66f4\u6ca1\u6d4b\uff0c\u56e0\u4e3a\u6211\u6ca1\u6709\u80fd\u5f00\u7684\u673a\u5668\uff0c\u4e0a\u9762\u7684\u663e\u5b58\u6570\u5b57\u5168\u662f\u522b\u4eba\u7684\uff0c\u51fa\u5904\u6211\u6807\u4e86\u3002\u60f3\u5224\u65ad\u5b83\u80fd\u4e0d\u80fd\u66ff\u6389\u4f60\u73b0\u5728\u7684\u65b9\u6848\uff0c\u53ea\u6709\u4e00\u4e2a\u529e\u6cd5\uff1a\u6311\u4e8c\u4e09\u5341\u6bb5\u8ddf\u4e1a\u52a1\u540c\u6b3e\u7684\u97f3\u9891\uff0c\u4e24\u8fb9\u5404\u8dd1\u4e00\u904d\uff0c\u81ea\u5df1\u542c\u3002\u8fd9\u4e2a\u9879\u76ee\u6700\u8ba9\u6211\u653e\u5fc3\u7684\u5730\u65b9\u4e5f\u5728\u8fd9\u513f\u2014\u2014\u5b83\u4e0d\u627f\u8bfa\u7ed3\u679c\uff0c\u53ea\u627f\u8bfa\u63a5\u53e3\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u4e09\u4e2a\u5c0f\u65f6\u7684\u4f1a\u8bae\u5f55\u97f3\uff0c\u6211\u539f\u6765\u90a3\u5957 whisper.cpp \u6d41\u6c34\u7ebf\u8981\u624b\u5de5\u6821\u4e24\u5c0f\u65f6\u3002FunASR \u5728 M2 \u4e0a\u8dd1\u901a\u4e86\uff1a5.55 \u79d2\u97f3\u9891 1.2 \u79d2\u51fa\u7ed3\u679c\uff0c\u5e26\u65f6\u95f4\u6233\u548c\u8bf4\u8bdd\u4eba\u3002\u4f46\u5b83\u4e0d\u662f\u88c5\u5b8c\u5c31\u80fd\u7528\u7684\u4e1c\u897f\u2014\u2014\u6807\u70b9\u6a21\u578b 1.1GB \u6bd4\u8bc6\u522b\u6a21\u578b\u8fd8\u5927\uff0c\u8bf4\u8bdd\u4eba\u5207\u6362\u70b9\u4e5f\u5f97\u9760\u6807\u70b9\u6a21\u578b\u6491\u7740\u3002<\/p>\n","protected":false},"author":1,"featured_media":10021510,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"_jetpack_memberships_contains_paid_content":false,"footnotes":""},"categories":[4],"tags":[],"class_list":["post-10021511","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-share"],"jetpack_sharing_enabled":true,"jetpack_featured_media_url":"https:\/\/itmax.cn\/wp-content\/uploads\/2026\/10\/funasr-cover.png","_links":{"self":[{"href":"https:\/\/itmax.cn\/index.php?rest_route=\/wp\/v2\/posts\/10021511","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/itmax.cn\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/itmax.cn\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/itmax.cn\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/itmax.cn\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=10021511"}],"version-history":[{"count":1,"href":"https:\/\/itmax.cn\/index.php?rest_route=\/wp\/v2\/posts\/10021511\/revisions"}],"predecessor-version":[{"id":10021512,"href":"https:\/\/itmax.cn\/index.php?rest_route=\/wp\/v2\/posts\/10021511\/revisions\/10021512"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/itmax.cn\/index.php?rest_route=\/wp\/v2\/media\/10021510"}],"wp:attachment":[{"href":"https:\/\/itmax.cn\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=10021511"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/itmax.cn\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=10021511"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/itmax.cn\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=10021511"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}