{"id":35909,"date":"2025-05-07T09:09:35","date_gmt":"2025-05-07T01:09:35","guid":{"rendered":"https:\/\/www.wsisp.com\/helps\/35909.html"},"modified":"2025-05-07T09:09:35","modified_gmt":"2025-05-07T01:09:35","slug":"%e5%9f%ba%e4%ba%8ewhisper%e6%b5%81%e5%bc%8f%e8%af%ad%e9%9f%b3%e8%af%86%e5%88%ab","status":"publish","type":"post","link":"https:\/\/www.wsisp.com\/helps\/35909.html","title":{"rendered":"\u57fa\u4e8ewhisper\u6d41\u5f0f\u8bed\u97f3\u8bc6\u522b"},"content":{"rendered":"

\u4e3a\u4e86\u5b9e\u73b0\u6301\u7eed\u76d1\u542c\u9ea6\u514b\u98ce\u5e76\u5728\u68c0\u6d4b\u5230\u58f0\u97f3\u65f6\u8fdb\u884c\u8f6c\u5f55，\u6211\u4eec\u53ef\u4ee5\u5c06\u6d41\u7684\u76d1\u542c\u65f6\u95f4\u8bbe\u7f6e\u4e3a\u65e0\u9650\u957f\u3002\u901a\u8fc7\u4f7f\u7528\u4e00\u4e2a\u97f3\u91cf\u95e8\u9650\u6765\u68c0\u6d4b\u662f\u5426\u6709\u58f0\u97f3，\u7136\u540e\u8fdb\u884c\u8f6c\u5f55\u3002<\/p>\n

\u5b89\u88c5\u4f9d\u8d56<\/h4>\n

\u786e\u4fdd\u5b89\u88c5\u5fc5\u8981\u7684\u5e93：<\/p>\n

pip install<\/span> torch torchaudio openai-whisper sounddevice numpy<\/p>\n

\u4ee3\u7801\u5b9e\u73b0<\/h4>\n

import<\/span> torch
\nimport<\/span> whisper
\nimport<\/span> sounddevice as<\/span> sd
\nimport<\/span> numpy as<\/span> np<\/p>\n

# \u52a0\u8f7d\u9884\u8bad\u7ec3\u7684 Whisper \u6a21\u578b<\/span>
\nmodel =<\/span> whisper.<\/span>load_model(<\/span>"base"<\/span>)<\/span><\/p>\n

# \u8bbe\u7f6e\u6a21\u578b\u4e3a\u8bc4\u4f30\u6a21\u5f0f<\/span>
\nmodel.<\/span>eval<\/span>(<\/span>)<\/span><\/p>\n

# \u5b9a\u4e49\u6d41\u5f0f\u89e3\u7801\u51fd\u6570<\/span>
\ndef<\/span> stream_decode<\/span>(<\/span>audio_buffer,<\/span> sample_rate=<\/span>16000<\/span>)<\/span>:<\/span>
\n audio_tensor =<\/span> torch.<\/span>tensor(<\/span>audio_buffer)<\/span>.<\/span>float<\/span>(<\/span>)<\/span>
\n result =<\/span> model.<\/span>transcribe(<\/span>audio_tensor,<\/span> fp16=<\/span>False<\/span>)<\/span>
\n return<\/span> result[<\/span>'text'<\/span>]<\/span><\/p>\n

# \u97f3\u9891\u7f13\u51b2\u533a\u548c\u5176\u4ed6\u53c2\u6570<\/span>
\nbuffer_size =<\/span> 16000<\/span> # \u6bcf\u4e2a\u97f3\u9891\u5757\u7684\u5927\u5c0f（1\u79d2）<\/span>
\naudio_buffer =<\/span> np.<\/span>zeros(<\/span>buffer_size *<\/span> 10<\/span>,<\/span> dtype=<\/span>np.<\/span>float32)<\/span> # \u9884\u755910\u79d2\u7f13\u51b2\u533a<\/span>
\nbuffer_offset =<\/span> 0<\/span>
\nsilence_threshold =<\/span> 0.01<\/span> # \u58f0\u97f3\u95e8\u9650<\/span><\/p>\n

# \u9ea6\u514b\u98ce\u56de\u8c03\u51fd\u6570<\/span>
\ndef<\/span> callback<\/span>(<\/span>indata,<\/span> frames,<\/span> time,<\/span> status)<\/span>:<\/span>
\n global<\/span> audio_buffer,<\/span> buffer_offset<\/p>\n

if<\/span> status:<\/span>
\n print<\/span>(<\/span>status,<\/span> flush=<\/span>True<\/span>)<\/span><\/p>\n

# \u8ba1\u7b97\u5f53\u524d\u97f3\u9891\u5757\u7684\u97f3\u91cf<\/span>
\n volume_norm =<\/span> np.<\/span>linalg.<\/span>norm(<\/span>indata)<\/span> *<\/span> 10<\/span>
\n if<\/span> volume_norm ><\/span> silence_threshold:<\/span>
\n # \u5c06\u65b0\u97f3\u9891\u6570\u636e\u590d\u5236\u5230\u7f13\u51b2\u533a<\/span>
\n audio_buffer[<\/span>buffer_offset:<\/span>buffer_offset+<\/span>frames]<\/span> =<\/span> indata[<\/span>:<\/span>,<\/span> 0<\/span>]<\/span>
\n buffer_offset +=<\/span> frames<\/p>\n

# \u5f53\u7f13\u51b2\u533a\u8fbe\u5230\u6216\u8d85\u8fc7\u8bbe\u5b9a\u7684\u5927\u5c0f\u65f6\u8fdb\u884c\u5904\u7406<\/span>
\n if<\/span> buffer_offset >=<\/span> buffer_size:<\/span>
\n text =<\/span> stream_decode(<\/span>audio_buffer[<\/span>:<\/span>buffer_size]<\/span>)<\/span>
\n print<\/span>(<\/span>f"Transcription: <\/span>{<\/span>text}<\/span><\/span>"<\/span><\/span>,<\/span> flush=<\/span>True<\/span>)<\/span><\/p>\n

# \u79fb\u52a8\u7f13\u51b2\u533a\u7684\u6570\u636e<\/span>
\n audio_buffer =<\/span> np.<\/span>roll(<\/span>audio_buffer,<\/span> –<\/span>buffer_size)<\/span>
\n buffer_offset -=<\/span> buffer_size
\n else<\/span>:<\/span>
\n # \u5982\u679c\u68c0\u6d4b\u5230\u7684\u97f3\u91cf\u4f4e\u4e8e\u95e8\u9650，\u5c06\u7f13\u51b2\u533a\u4f4d\u7f6e\u91cd\u7f6e<\/span>
\n buffer_offset =<\/span> 0<\/span><\/p>\n

# \u542f\u52a8\u9ea6\u514b\u98ce\u6d41<\/span>
\ndef<\/span> start_streaming<\/span>(<\/span>)<\/span>:<\/span>
\n stream =<\/span> sd.<\/span>InputStream(<\/span>callback=<\/span>callback,<\/span> channels=<\/span>1<\/span>,<\/span> samplerate=<\/span>16000<\/span>,<\/span> blocksize=<\/span>buffer_size)<\/span>
\n with<\/span> stream:<\/span>
\n print<\/span>(<\/span>"Listening…"<\/span>)<\/span>
\n while<\/span> True<\/span>:<\/span>
\n sd.<\/span>sleep(<\/span>1000<\/span>)<\/span> # \u7ee7\u7eed\u76d1\u542c<\/span><\/p>\n

# \u5f00\u59cb\u6d41\u5f0f\u89e3\u7801<\/span>
\nstart_streaming(<\/span>)<\/span><\/p>\n

\u4ee3\u7801\u8bf4\u660e<\/h4>\n
  • \n

    \u4f9d\u8d56\u52a0\u8f7d：<\/p>\n