# OpenAI Whisper 執行紀錄 ## 測試資料 - [財報狗176](https://podcasts.apple.com/tw/podcast/176-%E8%B2%A1%E7%B6%93%E6%99%82%E4%BA%8B%E6%94%BE%E5%A4%A7%E9%8F%A1-%E4%B8%AD%E5%9C%8B%E5%80%89%E4%BF%83%E8%A7%A3%E5%B0%81-%E4%BE%9B%E7%B5%A6%E7%9F%AD%E6%9C%9F%E5%85%A7%E5%9B%9E%E4%B8%8D%E4%BE%86/id1513810531?i=1000591039312) - 時長 27 分 52 秒 - 檔案大小 63.8 MB - [財報狗177](https://podcasts.apple.com/tw/podcast/177-%E9%81%94%E4%BA%BA%E8%81%8A%E6%8A%95%E8%B3%87-%E5%AE%B6%E6%97%8F%E8%BE%A6%E5%85%AC%E5%AE%A4%E6%8A%95%E8%B3%87%E9%95%B7%E8%81%8A%E6%8A%95%E8%B3%87%E4%B8%8A%E9%9B%86-%E9%87%91%E8%9E%8D%E9%A8%99%E5%B1%80%E5%85%AB%E5%8D%A6%E8%AB%87-ftx-%E6%83%A1%E8%A1%80-%E5%A6%82%E8%88%88/id1513810531?i=1000591262193) - 時長 75 分 - 檔案大小 172 MB - 用 large 模型處理花費 38 分鐘 ## 模型比較 | Model | Size | Execution Time | | -------- | -------- | -------- | | large | 2.87 GB | 13 分 59 秒 | | medium | 1.42 GB | 8 分 47 秒 | | small | 461 MB | 4 分 31 秒 | | base | 139 MB | 2 分 9 秒 | > For English-only applications, the .en models tend to perform better, especially for the tiny.en and base.en models. We observed that the difference becomes less significant for the small.en and medium.en models. ![](https://i.imgur.com/qQjfuDS.png) - 執行過程花費資源 | Model | VRAN | RAM | | -------- | -------- | -------- | | large | 11.21 GB | 3.23 GB | | medium | 6.25 GB | 1.67 GB | | small | 3.39 GB | 2.27 GB | | base | 2.32 GB | 2.54 GB | ## 程式碼 - [openai/whisper](https://github.com/openai/whisper) - [ArthurFDLR/whisper-youtube](https://github.com/ArthurFDLR/whisper-youtube) - 測試方式是用這份程式碼在 Google Colab 上執行 ## 各模型的輸出結果 這邊只有貼 txt ,其他還可以輸出 srt 、 vtt 字幕檔。 base 已經看到有錯字,所以沒有繼續測試 tiny ,目前還沒測試 `.en` 模型。 ## 結果比較 跟 large 輸出結果做比較,各自比較結果看左邊側欄連結。 | large | medium | | -------- | -------- | | 985 lines -514 Removals | 1159 lines +689 Additions | | large | small | | -------- | -------- | | 985 lines -573 Removals | 1176 lines +768 Additions | | large | base | | -------- | -------- | | 985 lines -707 Removals | 1246 lines +969 Additions | ## CPU https://github.com/openai/whisper/discussions/454