{"id":204055,"date":"2025-05-29T08:19:00","date_gmt":"2025-05-29T00:19:00","guid":{"rendered":"https:\/\/server.hk\/cnblog\/204055\/"},"modified":"2025-05-29T08:19:00","modified_gmt":"2025-05-29T00:19:00","slug":"%e9%80%82%e7%94%a8%e4%ba%8e%e6%82%a8%e7%9a%84%e5%ae%9e%e6%97%b6%e5%ba%94%e7%94%a8%e7%a8%8b%e5%ba%8f%e7%9a%84-supersonic-gpu-melspectrogram","status":"publish","type":"post","link":"https:\/\/server.hk\/cnblog\/204055\/","title":{"rendered":"\u9002\u7528\u4e8e\u60a8\u7684\u5b9e\u65f6\u5e94\u7528\u7a0b\u5e8f\u7684 Supersonic GPU MelSpectrogram"},"content":{"rendered":"

<\/b> <\/p>\n

\u9002\u7528\u4e8e\u60a8\u7684\u5b9e\u65f6\u5e94\u7528\u7a0b\u5e8f\u7684 Supersonic GPU MelSpectrogram<\/h1>\n

\u4eca\u65e5\u4e0d\u80af\u57cb\u5934\uff0c\u660e\u65e5\u4f55\u4ee5\u62ac\u5934\uff01\u6bcf\u65e5\u4e00\u53e5\u52aa\u529b\u81ea\u5df1\u7684\u8bdd\u54c8\u54c8~\u54c8\u55bd\uff0c\u4eca\u5929\u6211\u5c06\u7ed9\u5927\u5bb6\u5e26\u6765\u4e00\u7bc7\u300a\u9002\u7528\u4e8e\u60a8\u7684\u5b9e\u65f6\u5e94\u7528\u7a0b\u5e8f\u7684 Supersonic GPU MelSpectrogram\u300b<\/span>\uff0c\u4e3b\u8981\u5185\u5bb9\u662f\u8bb2\u89e3<\/span>\u7b49\u7b49\uff0c\u611f\u5174\u8da3\u7684\u670b\u53cb\u53ef\u4ee5\u6536\u85cf\u6216\u8005\u6709\u66f4\u597d\u7684\u5efa\u8bae\u5728\u8bc4\u8bba\u63d0\u51fa\uff0c\u6211\u90fd\u4f1a\u8ba4\u771f\u770b\u7684\uff01\u5927\u5bb6\u4e00\u8d77\u8fdb\u6b65\uff0c\u4e00\u8d77\u5b66\u4e60\uff01<\/p>\n

<\/p>\n

\u5728 simli\uff0c\u6211\u4eec\u6700\u5173\u5fc3\u7684\u662f\u5ef6\u8fdf\u3002\u6bd5\u7adf\uff0c\u8fd9\u5c31\u662f\u6211\u4eec\u7684\u76ee\u6807\uff1a\u4f4e\u5ef6\u8fdf\u89c6\u9891\u3002\u53e6\u4e00\u65b9\u9762\uff0c\u97f3\u9891\u673a\u5668\u5b66\u4e60\u4e2d\u4e00\u4e9b\u6700\u5e38\u7528\u7684\u7b97\u6cd5\u7684\u5b9e\u73b0\u901f\u5ea6\u975e\u5e38\u6162\u3002\u9700\u8981\u660e\u786e\u7684\u662f\uff0c\u8fd9\u4e9b\u5b9e\u73b0\u901a\u5e38\u9002\u5408\u521b\u5efa\u6a21\u578b\u672c\u8eab\u6216\u6279\u91cf\u63a8\u7406\u3002\u4f46\u5bf9\u4e8e simli \u7684\u6211\u4eec\u6765\u8bf4\uff0c\u51e0\u6beb\u79d2\u5c31\u53ef\u80fd\u610f\u5473\u7740\u89c6\u9891\u662f\u65ad\u65ad\u7eed\u7eed\u7684\u6df7\u4e71\u8fd8\u662f\u6d41\u7545\u3002
\u5bf9\u6211\u6765\u8bf4\u5e78\u8fd0\u7684\u662f\uff08\u4ee5\u53ca\u4f5c\u4e3a\u8bfb\u8005\u7684\u4ee3\u7406\uff09\uff0c\u672c\u6307\u5357\u4e0d\u9700\u8981\u592a\u591a\u6570\u5b66\u77e5\u8bc6\uff0c\u66f4\u806a\u660e\u7684\u4eba\u5df2\u7ecf\u5f04\u6e05\u695a\u5982\u4f55\u83b7\u5f97\u6b63\u786e\u7684\u7b54\u6848\uff0c\u6211\u4eec\u53ea\u662f\u8ba9\u8ba1\u7b97\u66f4\u52a0\u9ad8\u6548\u3002\u5982\u679c\u60a8\u9700\u8981\u66f4\u591a\u4fe1\u606f\u6765\u4e86\u89e3 melspectrogram \u5230\u5e95\u662f\u4ec0\u4e48\uff0c\u60a8\u53ef\u4ee5\u9605\u8bfb\u8fd9\u7bc7\u6587\u7ae0\u3002\u8ba1\u7b97\u9891\u8c31\u56fe\u7684\u65b9\u6cd5\u6709\u591a\u79cd\uff0c\u8fd9\u5728\u5f88\u5927\u7a0b\u5ea6\u4e0a\u53d6\u51b3\u4e8e\u60a8\u7684\u5e94\u7528\u7a0b\u5e8f\u3002\u56e0\u6b64\uff0c\u4e3a\u4e86\u65b9\u4fbf\u4f5c\u8005\uff0c\u6211\u4eec\u5c06\u91cd\u70b9\u653e\u5728\u8fd0\u884c\u5185\u90e8\u6a21\u578b\u6240\u9700\u7684\u6885\u5c14\u4e0a\u3002<\/p>\n

\u60a8\u5f88\u53ef\u80fd\u662f\u5728\u9047\u5230\u4f7f\u7528 librosa \u7684\u5b58\u50a8\u5e93\u540e\u6765\u5230\u8fd9\u91cc\u7684\u3002\u8001\u5b9e\u8bf4\uff0c\u8fd9\u662f\u4e00\u4e2a\u975e\u5e38\u65b9\u4fbf\u7684\u56fe\u4e66\u9986\u3002\u6709\u5927\u91cf\u5b9e\u7528\u7a0b\u5e8f\u3001\u8bfb\u53d6\u78c1\u76d8\u4e0a\u97f3\u9891\u7684\u7b80\u5355\u65b9\u6cd5\u4ee5\u53ca\u5feb\u901f\u8bbf\u95ee\u8bb8\u591a\u5e38\u7528\u529f\u80fd\uff08\u4f8b\u5982\u97f3\u9891\u91cd\u91c7\u6837\u3001\u901a\u9053\u7f29\u6df7\u7b49\uff09\u3002\u5728\u6211\u4eec\u7684\u4f8b\u5b50\u4e2d\uff0c\u6211\u4eec\u5bf9\u4e00\u79cd\u7279\u5b9a\u7684\u529f\u80fd\u611f\u5174\u8da3\uff1a\u6885\u5c14\u8c31\u56fe\u8ba1\u7b97\u3002\u5728 librosa \u4e2d\uff0c\u83b7\u53d6\u6885\u5c14\u5149\u8c31\u56fe\u975e\u5e38\u7b80\u5355\u3002<\/p>\n

import librosa\n\n# load in any audio to test\nsampleaudio, sr = librosa.load(\"sample.mp3\", sr=none) # sr=none means the original sampling rate\nspectrogram = librosa.feature.melspectrogram(\n    y=sampleaudio,\n    sr=sr,\n    n_fft=int(0.05 * sr),  # 50ms\n    hop_length=int(0.0125 * sr),  # 12.5ms\n    win_length=int(0.05 * sr),\n)\n<\/pre>\n
\u5f88\u7b80\u5355\uff0c\u5728 gcp g2 \u865a\u62df\u673a\u4e0a\u5e73\u5747\u9700\u8981 2 \u6beb\u79d2\u5de6\u53f3\u3002\u55ef\uff0c\u6709\u4e24\u4e2a\u4e3b\u8981\u95ee\u9898\uff1a<\/p>\n
    \n
  1. \u901a\u5e38\uff0c\u5728\u4f7f\u7528\u6df1\u5ea6\u5b66\u4e60\u6a21\u578b\u65f6\uff0c\u60a8\u9700\u8981\u5728 gpu \u4e0a\u8fd0\u884c\u6a21\u578b\u3002\u8fd9\u610f\u5473\u7740\u94fe\u7684\u4e00\u90e8\u5206\u5728 cpu \u4e0a\u8fd0\u884c\uff0c\u7136\u540e\u5c06\u7ed3\u679c\u590d\u5236\u56de gpu\u3002\u5bf9\u4e8e\u6279\u91cf\u63a8\u7406\uff0c\u8fd9\u57fa\u672c\u4e0a\u6ca1\u95ee\u9898\uff0c\u56e0\u4e3a\u60a8\u5e94\u8be5\u6536\u96c6 gpu\/\u4f20\u8f93\u4e0a\u80fd\u591f\u5bb9\u7eb3\u7684\u5c3d\u53ef\u80fd\u591a\u7684\u6570\u636e\u3002\u7136\u800c\uff0c\u5728\u6211\u4eec\u7684\u4f8b\u5b50\u4e2d\uff0c\u6211\u4eec\u7ecf\u5e38\u4e00\u6b21\u5904\u7406\u4e00\u5e27\uff0c\u4ee5\u51cf\u5c11\u7b49\u5f85\u548c\u5904\u7406\u65f6\u95f4\u3002<\/li>\n
  2. \u6211\u4eec\u7684\u603b\u65f6\u95f4\u9884\u7b97\u7ea6\u4e3a 33 \u6beb\u79d2\/\u5e27\u3002\u8fd9\u5305\u62ec\u4ece api \u670d\u52a1\u5668\u5230 ml \u63a8\u7406\u670d\u52a1\u5668\u7684\u4f20\u8f93\u5ef6\u8fdf\u3001cpu \u5230 gpu \u7684\u590d\u5236\u3001\u9884\u5904\u7406\u548c\u6a21\u578b\uff08\u5305\u62ec\u6885\u5c14\u8c31\u56fe\uff09\u7684\u540e\u5904\u7406\u3002\u5f53\u60a8\u7684\u9884\u7b97\u5982\u6b64\u7d27\u5f20\u65f6\uff0c\u6bcf\u4e00\u6beb\u79d2\u90fd\u5f88\u91cd\u8981\u3002\u8fd9\u4e24\u6beb\u79d2\u5b9e\u9645\u4e0a\u6709\u52a9\u4e8e\u4e3a simli \u63d0\u4f9b\u4e00\u4e2a\u53ef\u5de5\u4f5c\u7684\u5b9e\u65f6\u6e32\u67d3\u89c6\u9891\u6d41\uff08\u5f53\u7136\uff0c\u8fd9\u662f\u8bb8\u591a\u4f18\u5316\uff0c\u6bcf\u4e2a\u4f18\u5316\u90fd\u503c\u5f97\u4e00\u4e24\u6beb\u79d2\uff09\u3002<\/li>\n<\/ol>\n
    \u5728\u5c1d\u8bd5\u4e86\u89e3\u5176\u4ed6\u4eba\u662f\u5982\u4f55\u505a\u5230\u8fd9\u4e00\u70b9\u65f6\uff08\u5e78\u8fd0\u7684\u662f\uff0c\u8fd9\u5bf9\u6211\u4eec\u6765\u8bf4\u4e0d\u662f\u4e00\u4e2a\u72ec\u7279\u7684\u95ee\u9898\uff09\uff0c\u6211\u53d1\u73b0\u8fd9\u7bc7\u6587\u7ae0\u89e3\u91ca\u4e86\u6885\u5c14\u8c31\u56fe\u7684\u5de5\u4f5c\u539f\u7406\uff0c\u5e76\u63d0\u4f9b\u4e86\u4e00\u4e2a\u53c2\u8003\u5b9e\u73b0\uff0c\u7531\u4e8e\u67d0\u79cd\u539f\u56e0\uff0c\u8be5\u5b9e\u73b0\u4ec5\u82b1\u8d39\u4e86 1 \u6beb\u79d2\uff0850 \uff05 \u6539\u8fdb\uff09\u3002\u8fd9\u662f\u4e00\u4e2a\u597d\u7684\u5f00\u59cb\uff0c\u4f46\u4ecd\u7136\u5b58\u5728\u7b2c\u4e00\u4e2a\u95ee\u9898\uff0c\u5e76\u975e\u6240\u6709\u5185\u5bb9\u90fd\u5728 gpu \u4e0a\u3002\u6211\u4eec\u6b63\u5728\u4f7f\u7528 pytorch\uff0c\u5e76\u4e00\u76f4\u4f9d\u8d56 torch.compile \u548c mode=reduce-overhead \u6765\u6700\u5927\u7a0b\u5ea6\u5730\u63d0\u9ad8\u901f\u5ea6\u3002\u7136\u800c\uff0c\u50cf\u8fd9\u6837\u7684\u6570\u636e\u4f20\u8f93\u53ef\u80fd\u4f1a\u964d\u4f4e\u6027\u80fd\uff0c\u56e0\u4e3a pytorch \u7f16\u8bd1\u5668\u4e5f\u65e0\u6cd5\u4f18\u5316\u8be5\u51fd\u6570\u3002\u89e3\u51b3\u65b9\u6848\u6709\u70b9\u7e41\u7410\u4f46\u662f\u76f8\u5bf9\u7b80\u5355\uff0c\u7528torch\u91cd\u5199\u4e00\u4e0b\u5373\u53ef\u3002 pytorch \u56e2\u961f\u5df2\u786e\u4fdd\u5176\u8bb8\u591a\u8bed\u6cd5\u548c\u529f\u80fd\u5c3d\u53ef\u80fd\u63a5\u8fd1 numpy\uff08\u4e00\u4e9b\u8fb9\u7f18\u60c5\u51b5\u901a\u5e38\u90fd\u6709\u8be6\u7ec6\u8bb0\u5f55\uff0c\u9664\u4e86\u8ba9\u6211\u8ff7\u5931\u4e86\u51e0\u5929\u7684\u60c5\u51b5\uff0c\u4f46\u8fd9\u662f\u53e6\u4e00\u4e2a\u535a\u5ba2\u7684\u6545\u4e8b\uff09 .<\/p>\n
    \u56e0\u6b64\uff0c\u4e3a\u4e86\u6210\u529f\u91cd\u5199 pytorch \u4e2d\u7684\u6240\u6709\u5185\u5bb9\uff0c\u6211\u4eec\u9700\u8981\u6267\u884c\u51e0\u4e2a\u6b65\u9aa4\u3002\u6885\u5c14\u8c31\u56fe\u53ef\u4ee5\u5206\u4e3a\u4e09\u4e2a\u6b65\u9aa4\uff1a<\/p>\n
      \n
    • \u8ba1\u7b97\u77ed\u65f6\u5085\u91cc\u53f6\u53d8\u6362<\/li>\n
    • \u751f\u6210\u6885\u5c14\u97f3\u9636\u9891\u7387\u5e93<\/li>\n
    • \u751f\u6210\u9891\u8c31\u56fe\u3002<\/li>\n<\/ul>\n
      \u6709\u597d\u6d88\u606f\u4e5f\u6709\u574f\u6d88\u606f\u3002\u597d\u6d88\u606f\u662f\u6240\u6709\u5fc5\u9700\u7684\u529f\u80fd\u90fd\u53ef\u4ee5\u5728 pytorch \u6216 torchaudio \u4e2d\u8f7b\u677e\u83b7\u5f97\u3002\u574f\u6d88\u606f\u662f\u9ed8\u8ba4\u884c\u4e3a\u4e0e librosa \u6709\u5f88\u5927\u4e0d\u540c\uff0c\u56e0\u6b64\u9700\u8981\u8fdb\u884c\u5927\u91cf\u914d\u7f6e\u548c\u53cd\u590d\u8bd5\u9a8c\u624d\u80fd\u4f7f\u5176\u6b63\u786e\u3002\u6211\u7ecf\u5386\u8fc7\u8fd9\u4e00\u5207\uff0c\u6211\u5206\u4eab\u8fd9\u4e9b\u4fe1\u606f\u662f\u56e0\u4e3a\u6211\u4ec0\u81f3\u4e0d\u5e0c\u671b\u6211\u6700\u5927\u7684\u654c\u4eba\u906d\u53d7\u8fd9\u79cd\u5730\u72f1\u822c\u7684\u6253\u51fb\u3002\u6211\u4eec\u9700\u8981\u7406\u89e3\u7684\u4e00\u4ef6\u4e8b\u662f\uff0c\u8fd9\u6bb5\u4ee3\u7801\u4e25\u91cd\u4f9d\u8d56\u4e8e\u7f13\u5b58\u4e00\u4e9b\u7ed3\u679c\u4ee5\u4f9b\u4ee5\u540e\u4f7f\u7528\u3002\u8fd9\u662f\u5728\u9884\u751f\u6210\u6240\u6709\u9759\u6001\u6570\u7ec4\u7684\u521d\u59cb\u5316\u51fd\u6570\u4e2d\u5b8c\u6210\u7684\uff08\u4f8b\u5982\uff0c\u6885\u5c14\u9891\u7387\u5e93\u53d6\u51b3\u4e8e\u91c7\u6837\u7387\u548c\u6240\u9700\u7684\u6885\u5c14\u6570\u91cf\uff09\u3002\u8fd9\u662f\u6211\u4eec\u4f7f\u7528 pytorch \u4f18\u5316\u7684 melspectrogram \u51fd\u6570<\/p>\n
      \nimport torch\n\nif torch.cuda.is_available\n    @torch.compile(mode=\"reduce-overhead\")\nelse:\n    @torch.compile\ndef melspecrogram_torch(wav:torch.tensor,sample_rate:int, hann_window: torch.tensor, mel_basis: torch.tensor):\n    stftwav = torch.stft(\n            wav,\n            n_fft=int(sample_rate*0.05),\n            win_length=int(sample_rate*0.05),\n            hop_length=int(sample_rate*0.0125),\n            window=hann_window,\n            pad_mode=\"constant\",\n            return_complex=true,\n        ).abs()\n    stftwav = stftwav.squeeze()\n    mel_stftwav = torch.mm(mel_basis, stftwav)\n    return mel_stftwav\n\ndevice = \"cuda\" if torch.cuda.is_available() else \"cpu\"\n\nmelspectrogram_torch(\n    sampleaudio,\n    sr,\n    torch.hann_window(int(sample_rate*0.05), device=device, dtype=torch.float32),\n    torchaudio.functional.melscale_fbanks(\n        sample_rate=sr,\n        n_freqs=(int(sample_rate*0.05) \/\/ 2 + 1),\n        norm=\"slaney\", # this is the normalization algorithm used by librosa\n        # this is an example that's related to our own pipeline, check what you need for yours\n        n_mels=80,\n        f_min=55,\n        f_max=7600,\n    )\n    .t.to(device)\n)\n\n<\/pre>\n
      \u521d\u59cb\u7f16\u8bd1\u8fd0\u884c\u540e\uff0c\u6211\u4eec\u4f7f\u7528 nvidia l4 gpu\uff08\u7f13\u5b58 hann_window \u548c melscale_fbanks\uff09\u6d4b\u91cf\u8be5\u51fd\u6570\u9700\u8981 350 \u5fae\u79d2\u3002\u8c03\u6574\u540e\u7684\u8c03\u7528\u5c06\u5982\u4e0b\u6240\u793a\uff1a<\/p>\n
      hann=torch.hann_window(int(sample_rate*0.05), device=device, dtype=torch.float32),\nmelscale=torchaudio.functional.melscale_fbanks(\n        sample_rate=sr,\n        n_freqs=(int(sample_rate*0.05) \/\/ 2 + 1),\n        norm=\"slaney\", # this is the normalization algorithm used by librosa\n        # this is an example that's related to our own pipeline, check what you need for yours\n        n_mels=80,\n        f_min=55,\n        f_max=7600,\n    )\n    .T.to(device)\nmelspectrogram_torch(\n    sampleAudio,\n    sr,\n    hann,\n    melscale,\n)\n<\/pre>\n
      \u8fd9\u662f\u5173\u4e8e\u6211\u4eec\u5982\u4f55\u4f18\u5316\u90e8\u7f72\u7684\u9884\u8bad\u7ec3\u6a21\u578b\u3001\u4f18\u5316\u9884\u5904\u7406\u548c\u540e\u5904\u7406\u6b65\u9aa4\u7684\u4e00\u7cfb\u5217\u6587\u7ae0\u7684\u4e00\u90e8\u5206\u3002\u60a8\u53ef\u4ee5\u67e5\u770b https:\/\/www.simli.com\/demo \u67e5\u770b\u5df2\u90e8\u7f72\u7684\u6a21\u578b\u4ee5\u53ca\u6211\u4eec\u63d0\u4f9b\u7684\u6700\u4f4e\u5ef6\u8fdf\u7684\u5934\u50cf<\/p>\n
      \u5230\u8fd9\u91cc\uff0c\u6211\u4eec\u4e5f\u5c31\u8bb2\u5b8c\u4e86\u300a\u9002\u7528\u4e8e\u60a8\u7684\u5b9e\u65f6\u5e94\u7528\u7a0b\u5e8f\u7684 Supersonic GPU MelSpectrogram\u300b\u7684\u5185\u5bb9\u4e86\u3002\u4e2a\u4eba\u8ba4\u4e3a\uff0c\u57fa\u7840\u77e5\u8bc6\u7684\u5b66\u4e60\u548c\u5de9\u56fa\uff0c\u662f\u4e3a\u4e86\u66f4\u597d\u7684\u5c06\u5176\u8fd0\u7528\u5230\u9879\u76ee\u4e2d\uff0c\u6b22\u8fce\u5173\u6ce8\u516c\u4f17\u53f7\uff0c\u5e26\u4f60\u4e86\u89e3\u66f4\u591a\u5173\u4e8e\u7684\u77e5\u8bc6\u70b9\uff01<\/p>\n
            \u7248\u672c\u58f0\u660e \u672c\u6587\u8f6c\u8f7d\u4e8e\uff1adev.to \u5982\u6709\u4fb5\u72af\uff0c\u8bf7\u8054\u7cfb\u5220\u9664<\/p>\n","protected":false},"excerpt":{"rendered":"
      \u9002\u7528\u4e8e\u60a8\u7684\u5b9e\u65f6\u5e94\u7528\u7a0b\u5e8f\u7684 Sup...<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"closed","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[4925],"tags":[],"class_list":["post-204055","post","type-post","status-publish","format-standard","hentry","category-4925"],"_links":{"self":[{"href":"https:\/\/server.hk\/cnblog\/wp-json\/wp\/v2\/posts\/204055","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/server.hk\/cnblog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/server.hk\/cnblog\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/server.hk\/cnblog\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/server.hk\/cnblog\/wp-json\/wp\/v2\/comments?post=204055"}],"version-history":[{"count":0,"href":"https:\/\/server.hk\/cnblog\/wp-json\/wp\/v2\/posts\/204055\/revisions"}],"wp:attachment":[{"href":"https:\/\/server.hk\/cnblog\/wp-json\/wp\/v2\/media?parent=204055"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/server.hk\/cnblog\/wp-json\/wp\/v2\/categories?post=204055"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/server.hk\/cnblog\/wp-json\/wp\/v2\/tags?post=204055"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}