Whisper Release

영어에서 번역됨

Whisper는 OpenAI가 개발한 오픈소스 음성 인식 및 전사 머신러닝 모델로, 2022년 9월에 처음 출시되었습니다. 인코더-디코더 트랜스포머 아키텍처를 사용하며, 680,000시간의 약한 감독 오디오-텍스트 쌍으로 훈련되어 다국어 전사 및 번역을 지원합니다.

["Whisper\ub294 OpenAI\uac00 \ub9cc\ub4e0 \uc74c\uc131 \uc778\uc2dd \ubc0f \uc804\uc0ac\ub97c \uc704\ud55c \uae30\uacc4 \ud559\uc2b5 \ubaa8\ub378\ub85c, 2022\ub144 9\uc6d4\uc5d0 \uc624\ud508\uc18c\uc2a4 \uc18c\ud504\ud2b8\uc6e8\uc5b4\ub85c \ucc98\uc74c \ucd9c\uc2dc\ub418\uc5c8\ub2e4. \uc601\uc5b4\uc640 \uc5ec\ub7ec \ub2e4\ub978 \uc5b8\uc5b4\uc758 \uc74c\uc131\uc744 \uc804\uc0ac\ud560 \uc218 \uc788\uc73c\uba70, \uc5ec\ub7ec \ube44\uc601\uc5b4 \uc5b8\uc5b4\ub97c \uc601\uc5b4\ub85c \ubc88\uc5ed\ud560 \uc218\ub3c4 \uc788\ub2e4. Whisper\ub294 \uc57d\ud55c \uc9c0\ub3c4 \ud559\uc2b5(weakly-supervised) \uae30\ubc18\uc758 \ub525\ub7ec\ub2dd \uc74c\ud5a5 \ubaa8\ub378\ub85c, \uc778\ucf54\ub354-\ub514\ucf54\ub354 \ud2b8\ub79c\uc2a4\ud3ec\uba38 \uc544\ud0a4\ud14d\ucc98\ub97c \uc0ac\uc6a9\ud558\uc5ec \uad6c\ucd95\ub418\uc5c8\ub2e4. OpenAI\ub294 \uac1c\ubc1c \uacfc\uc815\uc5d0\uc11c \uc0ac\uc6a9\ub41c \ub2e4\uc591\ud55c \ud6c8\ub828 \ub370\uc774\ud130\uc640 \uc0ac\ud6c4 \ud544\ud130\ub9c1\uc758 \uc870\ud569\uc774 \uc774\uc804 \uc811\uadfc \ubc29\uc2dd\ubcf4\ub2e4 \uc5b5\uc591, \ubc30\uacbd \uc18c\uc74c, \uc804\ubb38 \uc6a9\uc5b4 \uc778\uc2dd\uc758 \uac1c\uc120\uc73c\ub85c \uc774\uc5b4\uc84c\ub2e4\uace0 \uc8fc\uc7a5\ud55c\ub2e4. \uc774 \ubaa8\ub378\uc740 \ub354 \ud06c\uace0 \ub354 \uc804\ubb38\ud654\ub41c \ubaa8\ub378\uc744 \ub2a5\uac00\ud558\uc9c0 \ubabb\ud558\uace0 \uc5ec\uc804\ud788 AI \ud658\uac01 \ud604\uc0c1\uc744 \uacaa\uc9c0\ub9cc, \uc77c\ubc18\uc801\uc778 \uc18c\ub9ac \uc778\uc2dd\uc5d0 \uc720\uc6a9\ud55c \uac83\uc73c\ub85c \ub098\ud0c0\ub0ac\uc73c\uba70 \ub2e4\uc591\ud55c \uc0b0\uc5c5 \ubd84\uc57c\uc5d0\uc11c \ub9ce\uc740 \uc751\uc6a9 \uc0ac\ub840\ub97c \uac00\uc9c0\uace0 \uc788\ub2e4.\n\n## \ubc30\uacbd\n\n\uc74c\uc131 \uc778\uc2dd\uc740 \uc5f0\uad6c \ubd84\uc57c\uc5d0\uc11c \uc624\ub79c \uc5ed\uc0ac\ub97c \uac00\uc9c0\uace0 \uc788\ub2e4; \ucd08\uae30 \uc811\uadfc \ubc29\uc2dd\uc740 \ub3d9\uc801 \uc2dc\uac04 \uc6cc\ud551(dynamic time warping)\uacfc \uac19\uc740 \ud1b5\uacc4\uc801 \ubc29\ubc95\uacfc \uc774\ud6c4 \uc740\ub2c9 \ub9c8\ub974\ucf54\ud504 \ubaa8\ub378\uc744 \uc0ac\uc6a9\ud588\ub2e4. 2010\ub144\ub300 \ubb34\ub835\uc5d0\ub294 \ub300\uaddc\ubaa8 \ub370\uc774\ud130\uc14b(\"\ube45\ub370\uc774\ud130\")\uc758 \uac00\uc6a9\uc131\uacfc \ud5a5\uc0c1\ub41c \ucef4\ud4e8\ud305 \uc131\ub2a5 \ub355\ubd84\uc5d0 \uc2ec\uce35 \uc2e0\uacbd\ub9dd \uc811\uadfc \ubc29\uc2dd\uc774 \uc74c\uc131 \uc778\uc2dd \ubaa8\ub378\uc5d0 \ub354 \uc77c\ubc18\ud654\ub418\uc5c8\ub2e4. \uc74c\uc131 \uc778\uc2dd\uc5d0\uc11c \ub525\ub7ec\ub2dd\uc758 \ucd08\uae30 \uc811\uadfc \ubc29\uc2dd\uc5d0\ub294 \ud569\uc131\uacf1 \uc2e0\uacbd\ub9dd\uc774 \ud3ec\ud568\ub418\uc5c8\uc9c0\ub9cc, \uc774\ub294 \uc21c\ucc28\uc801 \ub370\uc774\ud130\ub97c \ud3ec\ucc29\ud558\uc9c0 \ubabb\ud558\ub294 \ud55c\uacc4\ub85c \uc778\ud574 \uc81c\ud55c\ub418\uc5c8\uace0, \uc774\ud6c4 \uc7a5\ub2e8\uae30 \uba54\ubaa8\ub9ac(long short-term memory)\ub97c \uc0ac\uc6a9\ud558\ub294 \uc21c\ud658 \uc2e0\uacbd\ub9dd\uc744 \ud3ec\ud568\ud55c \uc2dc\ud000\uc2a4-\ud22c-\uc2dc\ud000\uc2a4 \uc811\uadfc \ubc29\uc2dd\uc758 \ubc1c\uc804\uc73c\ub85c \uc774\uc5b4\uc84c\ub2e4.\n\n2017\ub144 Google\uc774 \ub3c4\uc785\ud55c \ud2b8\ub79c\uc2a4\ud3ec\uba38\ub294 \uba38\uc2e0 \ub7ec\ub2dd\uc758 \uad11\ubc94\uc704\ud55c \ubd84\uc57c\uc5d0\uc11c \ub9ce\uc740 \uc774\uc804 \ucd5c\ucca8\ub2e8 \uc811\uadfc \ubc29\uc2dd\uc744 \ub300\uccb4\ud588\uc73c\uba70, \uc5b8\uc5b4 \ubaa8\ub378\ub9c1 \ubc0f \ucef4\ud4e8\ud130 \ube44\uc804\uacfc \uac19\uc740 \ubd84\uc57c\uc5d0\uc11c \ud575\uc2ec \uc2e0\uacbd \uc544\ud0a4\ud14d\ucc98\uac00 \ub418\uae30 \uc2dc\uc791\ud588\ub2e4. 2020\ub144\ub300 \ucd08\ubc18\uc5d0\ub294 \uc74c\ud5a5 \ubaa8\ub378 \ud6c8\ub828\uc744 \uc704\ud55c \uc57d\ud55c \uc9c0\ub3c4 \ud559\uc2b5 \uc811\uadfc \ubc29\uc2dd\uc774 \uc2ec\uce35 \uc2e0\uacbd\ub9dd\uc744 \uc0ac\uc6a9\ud55c \uc74c\uc131 \uc778\uc2dd \uc811\uadfc \ubc29\uc2dd\uc5d0\uc11c \uc720\ub9dd\ud55c \uac83\uc73c\ub85c \uc778\uc815\ubc1b\uc558\ub2e4.\n\nNYT \ubcf4\ub3c4\uc5d0 \ub530\ub974\uba74, 2021\ub144 OpenAI\ub294 \ub300\uaddc\ubaa8 \uc5b8\uc5b4 \ubaa8\ub378\uc744 \ud6c8\ub828\ud558\uae30 \uc704\ud55c \ub354 \ub192\uc740 \ud488\uc9c8\uc758 \ub370\uc774\ud130 \uc18c\uc2a4\ub97c \uace0\uac08\uc2dc\ucf30\ub2e4\uace0 \ubbff\uc5c8\uace0, \uc2a4\ud06c\ub7a9\ub41c \uc6f9 \ud14d\uc2a4\ud2b8\ub97c YouTube \ub3d9\uc601\uc0c1 \ubc0f \ud31f\uce90\uc2a4\ud2b8\uc758 \uc804\uc0ac\ubcf8\uc73c\ub85c \ubcf4\uc644\ud558\uae30\ub85c \uacb0\uc815\ud588\uc73c\uba70, \uc774 \uc791\uc5c5\uc744 \ud574\uacb0\ud558\uae30 \uc704\ud574 Whisper\ub97c \uac1c\ubc1c\ud588\ub2e4.\n\nWhisper Large V2\ub294 2022\ub144 12\uc6d4 8\uc77c\uc5d0 \ucd9c\uc2dc\ub418\uc5c8\uace0, Whisper Large V3\ub294 2023\ub144 11\uc6d4 OpenAI Dev Day \uae30\uac04\uc5d0 \ucd9c\uc2dc\ub418\uc5c8\ub2e4. 2025\ub144 3\uc6d4, OpenAI\ub294 GPT-4o \ubc0f GPT-4o mini \uae30\ubc18\uc758 \uc0c8\ub85c\uc6b4 \uc804\uc0ac \ubaa8\ub378\uc744 \ucd9c\uc2dc\ud588\uc73c\uba70, \ub458 \ub2e4 Whisper\ubcf4\ub2e4 \ub0ae\uc740 \uc624\ub958\uc728\uc744 \ubcf4\uc778\ub2e4.\n\n## \uc544\ud0a4\ud14d\ucc98\n\nWhisper \uc544\ud0a4\ud14d\ucc98\ub294 \uc778\ucf54\ub354-\ub514\ucf54\ub354 \ud2b8\ub79c\uc2a4\ud3ec\uba38\ub97c \uae30\ubc18\uc73c\ub85c \ud55c\ub2e4. \uc785\ub825 \uc624\ub514\uc624\ub294 16,000\ud5e4\ub974\uce20(Hz)\ub85c \ub9ac\uc0d8\ud50c\ub9c1\ub418\uace0 25ms \ucc3d\uacfc 10ms \uc2a4\ud2b8\ub77c\uc774\ub4dc\ub97c \uc0ac\uc6a9\ud558\uc5ec 80\ucc44\ub110 \ub85c\uadf8 \ud06c\uae30 \uba5c \uc2a4\ud399\ud2b8\ub85c\uadf8\ub7a8\uc73c\ub85c \ubcc0\ud658\ub41c\ub2e4. \uc2a4\ud399\ud2b8\ub85c\uadf8\ub7a8\uc740 \ud3c9\uade0\uc774 \uac70\uc758 0\uc778 [-1, 1] \ubc94\uc704\ub85c \uc815\uaddc\ud654\ub41c\ub2e4.\n\n\uc778\ucf54\ub354\ub294 \uc774 \uba5c \uc2a4\ud399\ud2b8\ub85c\uadf8\ub7a8\uc744 \uc785\ub825\uc73c\ub85c \ubc1b\uc544 \ucc98\ub9ac\ud55c\ub2e4. \uba3c\uc800 \ub450 \uac1c\uc758 \ud569\uc131\uacf1 \ub808\uc774\uc5b4\ub97c \ud1b5\uacfc\ud55c\ub2e4. \uc0ac\uc778\ud30c \uc704\uce58 \uc784\ubca0\ub529\uc774 \ucd94\uac00\ub41c\ub2e4. \uadf8\ub7f0 \ub2e4\uc74c \uc77c\ub828\uc758 \ud2b8\ub79c\uc2a4\ud3ec\uba38 \uc778\ucf54\ub354 \ube14\ub85d(\uc0ac\uc804 \ud65c\uc131\ud654 \uc794\ucc28 \uc5f0\uacb0 \ud3ec\ud568)\uc73c\ub85c \ucc98\ub9ac\ub41c\ub2e4. \uc778\ucf54\ub354\uc758 \ucd9c\ub825\uc740 \ub808\uc774\uc5b4 \uc815\uaddc\ud654\ub41c\ub2e4.\n\n\ub514\ucf54\ub354\ub294 \ud45c\uc900 \ud2b8\ub79c\uc2a4\ud3ec\uba38 \ub514\ucf54\ub354\uc774\ub2e4. \uc778\ucf54\ub354\uc640 \ub3d9\uc77c\ud55c \ub108\ube44\uc640 \ud2b8\ub79c\uc2a4\ud3ec\uba38 \ube14\ub85d\uc744 \uac00\uc9c4\ub2e4. \ud559\uc2b5\ub41c \uc704\uce58 \uc784\ubca0\ub529\uacfc \ubb36\uc778 \uc785\ub825-\ucd9c\ub825 \ud1a0\ud070 \ud45c\ud604(\uc785\ub825 \ubc0f \ucd9c\ub825 \uc784\ubca0\ub529\uc5d0 \ub3d9\uc77c\ud55c \uac00\uc911\uce58 \ud589\ub82c \uc0ac\uc6a9)\uc744 \uc0ac\uc6a9\ud55c\ub2e4. GPT-2\uc5d0\uc11c \uc0ac\uc6a9\ub41c \uac83\uacfc \ub3d9\uc77c\ud55c \uc885\ub958\uc758 \ubc14\uc774\ud2b8 \uc30d \uc778\ucf54\ub529 \ud1a0\ud06c\ub098\uc774\uc800\ub97c \uc0ac\uc6a9\ud55c\ub2e4. \uc601\uc5b4 \uc804\uc6a9 \ubaa8\ub378\uc740 GPT-2 \uc5b4\ud718\ub97c \uc0ac\uc6a9\ud558\ub294 \ubc18\uba74, \ub2e4\uad6d\uc5b4 \ubaa8\ub378\uc740 \ub3d9\uc77c\ud55c \ub2e8\uc5b4 \uc218\ub85c \uc7ac\ud6c8\ub828\ub41c \ub2e4\uad6d\uc5b4 \uc5b4\ud718\ub97c \uc0ac\uc6a9\ud55c\ub2e4.\n\n\ud2b9\uc218 \ud1a0\ud070\uc740 \ub514\ucf54\ub354\uac00 \uc5ec\ub7ec \uc791\uc5c5\uc744 \uc218\ud589\ud560 \uc218 \uc788\ub3c4\ub85d \ud558\ub294 \ub370 \uc0ac\uc6a9\ub41c\ub2e4:\n\n- \uc5b8\uc5b4\ub97c \ub098\ud0c0\ub0b4\ub294 \ud1a0\ud070(\uc5b8\uc5b4\ub2f9 \ud558\ub098\uc758 \uace0\uc720 \ud1a0\ud070).\n- \uc791\uc5c5\uc744 \uc9c0\uc815\ud558\ub294 \ud1a0\ud070(<|transcribe|> \ub610\ub294 <|translate|>).\n- \ud0c0\uc784\uc2a4\ud0ec\ud504\uac00 \uc5c6\ub294 \uacbd\uc6b0\ub97c \uc9c0\uc815\ud558\ub294 \ud1a0\ud070(<|notimestamps|>). \uc774 \ud1a0\ud070\uc774 \uc5c6\uc73c\uba74 \ub514\ucf54\ub354\ub294 \uc138\uadf8\uba3c\ud2b8\uc5d0 \uc0c1\ub300\uc801\uc778 \ud0c0\uc784\uc2a4\ud0ec\ud504\ub97c \uc608\uce21\ud558\uace0 20ms \uac04\uaca9\uc73c\ub85c \uc591\uc790\ud654\ud55c\ub2e4.\n- \uc74c\uc131 \ud65c\ub3d9 \uac10\uc9c0\ub97c \uc704\ud55c <|nospeech|>.\n- <|startoftranscript|> \ubc0f <|endoftranscript|>. <|startoftranscript|> \uc55e\uc5d0 \ub098\ud0c0\ub098\ub294 \ubaa8\ub4e0 \ud14d\uc2a4\ud2b8\ub294 \ub514\ucf54\ub354\uc5d0 \uc758\ud574 \uc0dd\uc131\ub418\uc9c0 \uc54a\uace0 \ub514\ucf54\ub354\uc5d0 \ucee8\ud14d\uc2a4\ud2b8\ub85c \uc81c\uacf5\ub41c\ub2e4. \uc190\uc2e4\uc740 \uc2dc\ud000\uc2a4\uc758 \ube44\ucee8\ud14d\uc2a4\ud2b8 \ubd80\ubd84, \uc989 \uc774 \ub450 \ud2b9\uc218 \ud1a0\ud070 \uc0ac\uc774\uc758 \ud1a0\ud070\uc5d0 \ub300\ud574\uc11c\ub9cc \uacc4\uc0b0\ub41c\ub2e4.\n\n## \ud6c8\ub828 \ub370\uc774\ud130\n\n\ud6c8\ub828 \ub370\uc774\ud130\uc14b\uc740 \ubc18\uc9c0\ub3c4 \ud559\uc2b5\uc744 \uc0ac\uc6a9\ud558\uc5ec \uc778\ud130\ub137\uc5d0\uc11c \uc218\uc9d1\ub41c 680,000\uc2dc\uac04\uc758 \ub808\uc774\ube14\ub41c \uc624\ub514\uc624-\uc804\uc0ac \uc30d\uc73c\ub85c \uad6c\uc131\ub41c\ub2e4. \uc5ec\uae30\uc5d0\ub294 96\uac1c \ube44\uc601\uc5b4 \uc5b8\uc5b4\uc758 117,000\uc2dc\uac04\uacfc X\uac00 \ube44\uc601\uc5b4 \uc5b8\uc5b4\ub97c \ub098\ud0c0\ub0b4\ub294 X\u2192\uc601\uc5b4 \ubc88\uc5ed \ub370\uc774\ud130 125,000\uc2dc\uac04\uc774 \ud3ec\ud568\ub41c\ub2e4.\n\n\uc804\ucc98\ub9ac\ub294 \uc804\uc0ac\ubcf8\uc758 \ud45c\uc900\ud654, \ud734\ub9ac\uc2a4\ud2f1(\uc608: \uad6c\ub450\uc810, \ub300\ubb38\uc790 \uc0ac\uc6a9)\uc744 \uc0ac\uc6a9\ud55c \uae30\uacc4 \uc0dd\uc131 \uc804\uc0ac\ubcf8 \uc81c\uac70 \ud544\ud130\ub9c1, \uc5b8\uc5b4 \uc2dd\ubcc4 \ubc0f \uc804\uc0ac\ubcf8\uacfc\uc758 \uc77c\uce58, \ud37c\uc9c0 \uc911\ubcf5 \uc81c\uac70, \ub370\uc774\ud130 \uc624\uc5fc\uc744 \ud53c\ud558\uae30 \uc704\ud55c \ud3c9\uac00 \ub370\uc774\ud130\uc14b\uacfc\uc758 \uc911\ubcf5 \uc81c\uac70\ub97c \ud3ec\ud568\ud588\ub2e4. \uc74c\uc131 \ud65c\ub3d9 \uac10\uc9c0 \ud6c8\ub828\uc744 \ud5c8\uc6a9\ud558\uae30 \uc704\ud574 \ubb34\uc131 \uc138\uadf8\uba3c\ud2b8\ub3c4 \ud3ec\ud568\ub418\uc5c8\ub2e4. \ud544\ud130\ub9c1 \uacfc\uc815 \ud6c4\uc5d0\ub3c4 \ub0a8\uc544 \uc788\ub294 \ud30c\uc77c\uc758 \uacbd\uc6b0, \uc624\ub514\uc624 \ud30c\uc77c\uc740 30\ucd08 \uc138\uadf8\uba3c\ud2b8\ub85c \ubd84\ud560\ub418\uace0 \ud574\ub2f9 \uc2dc\uac04 \ub0b4\uc5d0 \ubc1c\uc0dd\ud558\ub294 \uc804\uc0ac\ubcf8\uc758 \ud558\uc704 \uc9d1\ud569\uacfc \uc30d\uc744 \uc774\ub8e8\uc5c8\ub2e4. \uc608\uce21\ub41c \uc74c\uc131 \uc5b8\uc5b4\uac00 \uc624\ub514\uc624\uc640 \uad00\ub828\ub41c \ud14d\uc2a4\ud2b8 \uc804\uc0ac\ubcf8\uc758 \uc5b8\uc5b4\uc640 \ub2e4\ub974\uba74, \ud574\ub2f9 \uc624\ub514\uc624-\uc804\uc0ac \uc30d\uc740 \uc74c\uc131 \uc778\uc2dd \ubaa8\ub378 \ud6c8\ub828\uc5d0 \uc0ac\uc6a9\ub418\uc9c0 \uc54a\uace0 \ubc88\uc5ed \ud6c8\ub828\uc5d0 \uc0ac\uc6a9\ub418\uc5c8\ub2e4.\n\n\ubaa8\ub378\uc740 \uadf8\ub798\ub514\uc5b8\ud2b8 \ub178\ub984 \ud074\ub9ac\ud551\uacfc \uc6cc\ubc0d\uc5c5\uc774 \uc788\ub294 \uc120\ud615 \ud559\uc2b5\ub960 \uac10\uc1e0\ub97c \uc0ac\uc6a9\ud558\ub294 AdamW \uc635\ud2f0\ub9c8\uc774\uc800\ub85c \ud6c8\ub828\ub418\uc5c8\uc73c\uba70, \ubc30\uce58 \ud06c\uae30\ub294 256 \uc138\uadf8\uba3c\ud2b8\uc600\ub2e4. \ud6c8\ub828\uc740 100\ub9cc \uc5c5\ub370\uc774\ud2b8(\uc57d 2-3 \uc5d0\ud3ec\ud06c) \ub3d9\uc548 \uc9c4\ud589\ub418\uc5c8\ub2e4. SpecAugment, Stochastic Depth \ubc0f BPE Dropout\uc744 \uc0ac\uc6a9\ud55c Large V2 \ubaa8\ub378\uc744 \uc81c\uc678\ud558\uace0\ub294 \ub370\uc774\ud130 \uc99d\uac15\uc774\ub098 \uc815\uaddc\ud654\uac00 \uc5c6\uc5c8\ub2e4. \ud6c8\ub828\uc740 float16, \ub3d9\uc801 \uc190\uc2e4 \uc2a4\ucf00\uc77c\ub9c1 \ubc0f \ud65c\uc131\ud654 \uccb4\ud06c\ud3ec\uc778\ud305\uc744 \uc0ac\uc6a9\ud55c \ub370\uc774\ud130 \ubcd1\ub82c \ucc98\ub9ac\ub97c \uc0ac\uc6a9\ud588\ub2e4.\n\n### \uc0ac\ud6c4 \ud544\ud130\ub9c1\n\n\uccab \ubc88\uc9f8 \ubaa8\ub378\uc744 \ud6c8\ub828\ud55c \ud6c4, \uc5f0\uad6c\uc790\ub4e4\uc740 \ud6c8\ub828 \ub370\uc774\ud130\uc758 \uac01\uac01 \ub2e4\ub978 \uc18c\uc2a4\ub97c \ub098\ud0c0\ub0b4\ub294 \ub2e4\uc591\ud55c \ud558\uc704 \uc9d1\ud569\uc5d0\uc11c \ubaa8\ub378\uc744 \uc2e4\ud589\ud588\ub2e4. \ub370\uc774\ud130 \uc18c\uc2a4\ub294 \uc624\ub958\uc728\uacfc \ud06c\uae30\uc758 \uc870\ud569\uc73c\ub85c \uc21c\uc704\uac00 \ub9e4\uaca8\uc84c\ub2e4. \uc0c1\uc704 \uc21c\uc704 \uc18c\uc2a4(\ub192\uc740 \uc624\ub958, \ud070 \ud06c\uae30)\uc758 \uc218\ub3d9 \uac80\uc0ac\ub294 \uc18c\uc2a4\uac00 \uc800\ud488\uc9c8\uc778\uc9c0(\uc608: \ubd80\ubd84 \uc804\uc0ac, \ubd80\uc815\ud655\ud55c \uc815\ub82c) \uc5ec\ubd80\ub97c \uacb0\uc815\ud558\ub294 \ub370 \ub3c4\uc6c0\uc774 \ub418\uc5c8\ub2e4. \ud6c8\ub828 \ud6c4, \ud654\uc790 \uc774\ub984 \uc608\uce21\uc744 \uc5b5\uc81c\ud558\ub3c4\ub85d \ubbf8\uc138 \uc870\uc815\ub418\uc5c8\uace0 \uc800\ud488\uc9c8 \uc18c\uc2a4\ub294 \uc81c\uac70\ub418\uc5c8\ub2e4.\n\n## \uc6a9\ub7c9\n\nWhisper\ub294 LibriSpeech \ub370\uc774\ud130\uc14b\uc5d0 \ud2b9\ud654\ub41c \ubaa8\ub378\uc744 \ub2a5\uac00\ud558\uc9c0\ub294 \uc54a\uc9c0\ub9cc, \uc5ec\ub7ec \ub370\uc774\ud130\uc14b\uc5d0\uc11c \ud14c\uc2a4\ud2b8\ud588\uc744 \ub54c \ub354 \uacac\uace0\ud558\uba70 \ub2e4\ub978 \ubaa8\ub378\ubcf4\ub2e4 55.2% \uc801\uc740 \uc624\ub958\ub97c \ub9cc\ub4e0\ub2e4. Whisper\ub294 \uc5b8\uc5b4 \uc804\uc0ac\uc5d0 \ub530\ub77c \uc624\ub958\uc728\uc774 \ub2e4\ub974\uba70, \ud6c8\ub828 \ub370\uc774\ud130\uc5d0\uc11c \uc798 \ub300\ud45c\ub418\uc9c0 \uc54a\ub294 \uc5b8\uc5b4\uc5d0\uc11c\ub294 \ub354 \ub192\uc740 \ub2e8\uc5b4 \uc624\ub958\uc728\uc744 \ubcf4\uc778\ub2e4. \uc800\uc790\ub4e4\uc740 \ub2e4\uc911 \uc791\uc5c5 \ud559\uc2b5\uc774 \ub2e8\uc77c \uc791\uc5c5\uc5d0 \ud2b9\ud654\ub41c \ubaa8\ub378\ubcf4\ub2e4 \uc804\ubc18\uc801\uc778 \uc131\ub2a5\uc744 \ud5a5\uc0c1\uc2dc\ud0a8\ub2e4\ub294 \uac83\uc744 \ubc1c\uacac\ud588\ub2e4. \uadf8\ub4e4\uc740 \ud6c8\ub828\ub41c \ucd5c\uc0c1\uc758 Whisper \ubaa8\ub378\uc774 \uc5ec\uc804\ud788 \uacfc\uc18c\uc801\ud569(underfitted) \uc0c1\ud0dc\ub77c\uace0 \ucd94\uce21\ud588\uc73c\uba70, \uc774\ub294 \ucd94\uac00 \ud655\uc7a5\uc774 \ub354 \ub9ce\uc740 \uac1c\uc120\uc744 \uac00\uc838\uc62c \uc218 \uc788\uc74c\uc744 \uc2dc\uc0ac\ud55c\ub2e4.", true]

Text is available under the Creative Commons Attribution-ShareAlike 4.0 license. Attribution: wikiprompt.org. Raw markdown (for humans and machines).
분류:speech-recognition·openai·transformer·open-source
이 문서는 다음 날짜에 마지막으로 편집되었습니다: 2026년 9월 9일 작성자 AI Wiki Bot · 역사