
ãç¡æã§æåèµ·ãããWhisperã®å°å ¥æ¹æ³ãšãæå€ãšç¥ãããŠããªã"SRTãã¡ã€ã«"ã®äŸ¿å©ãªæŽ»çšæ³
ã¯ããã«
ã¿ãªãããäŒè°ããã¬ãŒã³ã®æåèµ·ãããã©ãããŠããŸããïŒ
ææã®æåèµ·ãããµãŒãã¹ã¯ãããããããŸããã䜿ãé »åºŠãå€ããšã³ã¹ããããã¿ãŸããããå®ã¯ãOpenAIãå ¬éããŠãã WhisperïŒãŠã£ã¹ããŒïŒ ãšããé³å£°èªèã¢ãã«ãèªåã®PCã«å°å ¥ããã°ãå®å šç¡æã§æåèµ·ãããã§ãããã§ãã
ãã ããæ£çŽã«ãäŒããããšãã»ããã¢ããã«ã¯ã¡ãã£ãšããæéãããããŸããGPUããã£ãã»ããå¿«é©ã§ãããç°å¢æ§ç¯ãããè€éã§ãã
ããã§ããäžåºŠã»ããã¢ããããŠããŸãã°ãããšã¯è¿œå æéãªãã§äœåºŠã§ã䜿ããã®ã§ãã³ã¹ãã¯æåŒ·ã§ãã
ãã®èšäºã§ã¯ãWhisperã®å°å ¥æé ãšãç§ãæè¿èŠã€ãã SRTãã¡ã€ã«ã䜿ã£ã䟿å©ãªæŽ»çšæ³ ãã玹ä»ããŸãã
Whisperãšã¯ïŒ
Whisperã¯OpenAIãéçºã»å ¬éããŠããé³å£°èªèã¢ãã«ã§ããå€èšèªå¯Ÿå¿ã§ãæ¥æ¬èªã®èªè粟床ãããªãé«ãã§ãã
ãã€ã³ãããŸãšãããšããããªæãã§ãã
ç¡æã§äœ¿ããïŒããŒã«ã«å®è¡ã®å ŽåïŒ
æ¥æ¬èªãå«ãå€èšèªã«å¯Ÿå¿
é·æéã®é³å£°ããŸãšããŠåŠçã§ãã
ããã¹ãã ãã§ãªããåå¹ãã¡ã€ã«ãªã©è€æ°ã®åœ¢åŒã§åºåã§ãã
Windows PCãžã®å°å ¥æé
Whisperã¯äŸåé¢ä¿ãã©ã€ãã©ãªã®ããŒãžã§ã³ã«ææãªã®ã§ãå°çšã®ä»®æ³ç°å¢ãäœãããšã匷ãããããããŸããä»ã®ã©ã€ãã©ãªã远å ãããšåããªããªãããšãããã®ã§ãããã¯å€§äºãªãã€ã³ãã§ãã
ã¹ããã1ïŒäºåæºå
以äžã®ããŒã«ãå ã«ã€ã³ã¹ããŒã«ããŠãããŸãããã
â PythonïŒããŒãžã§ã³3.10ã3.12ãæšå¥šïŒ
Whisperèªäœã¯Python 3.8以äžã«å¯Ÿå¿ããŠããŸãããææ°ã®PyTorchïŒ2.10以éïŒãPython 3.10以äžãå¿
èŠãšãããããPython 3.10ã3.12ããããéžãã§ããã®ãç¡é£ã§ããäžéšã®ç°å¢ã§ã¯Python 3.13ã§åäœããªãã±ãŒã¹ãå ±åãããŠããã®ã§ãè¿·ã£ãã3.10ã3.12ãéžã³ãŸãããã
â¡ GPUé¢é£ããŒã«ïŒGPUãããå ŽåïŒ
GPUãæŽ»çšãããšæåèµ·ãããå§åçã«éããªããŸãã以äžãå¿
èŠã§ãã
Build Tools for Visual Studio 2022
NVIDIAãã©ã€ã
NVIDIA CUDAããŒã«ããã
NVIDIA cuDNN
CUDAãšcuDNNã®ããŒãžã§ã³ã¯ãPyTorchã®å¯Ÿå¿ç¶æ³ã«åãããŠéžã¶å¿ èŠããããŸããPyTorchã®å ¬åŒãµã€ãã§ãçŸåšãµããŒããããŠããCUDAããŒãžã§ã³ãå¿ ã確èªããŠãã ããã 2026幎3ææç¹ã§ã¯ãCUDA 12.6 / 12.8 / 13.0ãªã©ã察å¿ããŠããŸãã
GPUããªããŠãCPUã§åããŸãããåŠçæéãããªãé·ããªããŸããGPUãããPCããæã¡ãªãããã²æŽ»çšããŸãããã
⢠FFmpeg
åç»ãé³å£°ãã¡ã€ã«ã®å€æã«äœ¿ãããŒã«ã§ããWhisperãå
éšçã«å©çšããã®ã§å¿
é ã§ãã
FFmpegã®å ¬åŒãµã€ãããWindowsçãããŠã³ããŒã
`.7z`ãã¡ã€ã«ãå±éããŠã`C:\ffmpeg`ã«ã³ããŒ
ã·ã¹ãã ã®ç°å¢å€æ°Pathã« `C:\ffmpeg\bin` ã远å
ã³ãã³ãããã³ããã§ `where ffmpeg` ãšå ¥åããŠããã¹ã衚瀺ãããã°OK
ð¡ ãã€ã³ãïŒ ãã®FFmpegã¯ãåŸè¿°ããSRTãã¡ã€ã«ã®æŽ»çšã§ã倧掻èºããã®ã§ããã£ããå ¥ããŠãããŠãã ããã
ã¹ããã2ïŒPyTorchã®ã€ã³ã¹ããŒã«
PyTorchã®å ¬åŒãµã€ãã«ã¢ã¯ã»ã¹ããèªåã®ç°å¢ïŒOSãCUDA ããŒãžã§ã³ãªã©ïŒã«åã£ãã€ã³ã¹ããŒã«ã³ãã³ãã確èªããŠå®è¡ããŸãã
ã€ã³ã¹ããŒã«åŸã以äžã®ã³ãã³ãã§ããŒãžã§ã³ãšGPUã®å©çšå¯åŠã確èªã§ããŸãã
python -c "import torch; print(torch.__version__, torch.cuda.is_available())"`True` ãšè¡šç€ºãããã°ãGPUã䜿ããç¶æ ã§ãã
ã¹ããã3ïŒWhisperã®ã€ã³ã¹ããŒã«
以äžã®ã³ãã³ãã§Whisperãã€ã³ã¹ããŒã«ããŸãã
pip install -U openai-whisperææ°ã®éçºçãå ¥ãããå Žåã¯ã以äžã®ã³ãã³ãã§GitHubããçŽæ¥ã€ã³ã¹ããŒã«ããããšãã§ããŸãã
pip install git+https://github.com/openai/whisper.gitããã§å°å ¥ã¯å®äºã§ãã
ð¡ è£è¶³ïŒ Whisperã«ã¯tiny / base / small / medium / large / turboãªã©è€æ°ã®ã¢ãã«ãµã€ãºããããŸããturbo㯠large-v3 ãããŒã¹ã«é«éåãããã¢ãã«ã§ãè±èªã®æåèµ·ããã«ãããŠéåžžã«é«éã«åäœããŸãã粟床äœäžã¯æ¯èŒçå°ãªããšãããŠããŸãããçšéãé³å£°ã«ãã£ãŠã¯å·®ãåºãããã粟床éèŠã®å Žå㯠large ã medium ã®äœ¿çšãæ€èšãããšããã§ãããããªããturboã¯ç¿»èš³ã¿ã¹ã¯åãã«ã¯æé©åãããŠããªããããæ¥æ¬èªé³å£°ãè±èªã«ç¿»èš³ããçšéã§ã¯ multilingual ã¢ãã«ïŒmedium / large ãªã©ïŒã䜿ãã®ãæšå¥šãããŸãã

Whisperã®åºåãã¡ã€ã«åœ¢åŒãçè§£ãã
Whisperã§æåèµ·ãããå®è¡ãããšãããã¹ããã¡ã€ã«ã ãã§ãªããè€æ°ã®åœ¢åŒã§åºåã§ããã®ã倧ããªç¹åŸŽã§ãã
ãããçè§£ããŠãããšã掻çšã®å¹ ãã°ããšåºãããŸãã
TXTïŒã·ã³ãã«ãªå šæããã¹ã
æéæ å ±ãªãã®ãçŽç²ãªããã¹ãã§ããè°äºé²ã®ãããå°ãäœã£ãããChatGPTãªã©ã®AIã«èŠçŽããããã翻蚳çšã®çŽ æã«ãããããã®ã«æé©ã§ãã
ãšã«ãããäœã話ããããã ãåããã°ããããšããå Žé¢ã§ã¯ããã§ååã§ãã
SRTïŒåç»åå¹ã®å®çªãã©ãŒããã
ååå¹ã«çªå·ãä»ããéå§æå» â çµäºæå» â åå¹ããã¹ã ãšããæ§é ã«ãªã£ãŠããŸããåç»ç·šéãœãããYouTubeã®åå¹ã«ãã®ãŸãŸäœ¿ãããæãæ±çšæ§ã®é«ãåå¹åœ¢åŒã§ãã
ãã®èšäºã®åŸåã§ãSRTãã¡ã€ã«ã®ãåå¹ä»¥å€ãã®æŽ»çšæ³ã玹ä»ããŸãã ãããä»åã®æ¬é¡ã§ãã
VTTïŒWebåç»åãã®åå¹
SRTã«è¿ãã§ããããã©ãŠã¶ãHTML5åç»ãã¬ã€ã€ãŒãšã®çžæ§ããã圢åŒã§ããWebäžã§åå¹ã衚瀺ãããå Žåã¯ãã¡ããåããŠããŸãã
TSVïŒè¡šåœ¢åŒã®ããŒã¿
ã¿ãåºåãã§ãéå§æå»ã»çµäºæå»ã»ããã¹ããã䞊ã¶åœ¢åŒã§ããExcelãGoogleã¹ãã¬ããã·ãŒãã«è²Œãä»ãããããããŒã¿ãšããŠæŽçã»åæãããå Žåã«éå®ããŸãã
JSONïŒããã°ã©ã åŠçåãã®æ§é åããŒã¿
å šæããã¹ããã»ã°ã¡ã³ãããšã®æéæ å ±ãæ€åºèšèªãªã©ãæãæ å ±éãå€ã圢åŒã§ããèªäœã¢ããªãšã®é£æºããèªååã¯ãŒã¯ãããŒã«çµã¿èŸŒãã®ã«åããŠããŸãã
ããããã®äœ¿ãåã
| çšé | ããããåœ¢åŒ |
|------|-------------|
| è°äºé²ãäœããã | TXT ïŒ JSON |
| åç»ã«åå¹ãä»ããã | SRT ïŒ VTT |
| ããŒã¿ãšããŠåæããã | TSV ïŒ JSON |
| è¿·ã£ãã | `all`ïŒå
šåœ¢åŒåºåïŒ |
è¿·ã£ãã `all` ãæå®ããŠå šéšåºããŠããã®ãããããã§ããããšããå¿ èŠãªåœ¢åŒã ã䜿ãã°ããã®ã§ãæã¯ãããŸããã

ãæ¬é¡ãSRTãã¡ã€ã« à FFmpegã§åºããæŽ»çšæ³
ãããããä»åãã¡ã°ããäŒããããã£ãå 容ã§ãã
ç§ã¯ãããŸã§Whisperããã£ã±ããäŒè°ã®æåèµ·ããâTXTã§è°äºé²åããšããäœ¿ãæ¹ãããŠããŸãããã§ãæè¿ãSRTãã¡ã€ã«ã®ã¿ã€ã ã¹ã¿ã³ãæ å ±ã«æ³šç®ããããšã§ã掻çšã®å¹ ãäžæ°ã«åºãããŸããã
SRTãã¡ã€ã«ã®äžèº«ïŒã€ã¡ãŒãžïŒ
SRTãã¡ã€ã«ã®äžèº«ã¯ããããªæ§é ã«ãªã£ãŠããŸãã
1
00:00:05,000 --> 00:00:12,000
ããã§ã¯æ¬æ¥ã®ãã¬ãŒã³ãå§ããŸã
2
00:00:12,500 --> 00:00:20,000
ãŸãæåã«ã仿ã®å£²äžã«ã€ããŠãå ±åããŸã
3
00:00:20,500 --> 00:00:35,000
ãã¡ãã®ã°ã©ããã芧ãã ãã...ã€ãŸããããã€ãäœã話ãããããæéä»ãã§èšé²ãããŠããããã§ãã
掻çšäŸïŒãã¬ãŒã³ç·Žç¿ã®æ¯ãè¿ã
ããšãã°ããã¬ãŒã³ã®ç·Žç¿ããããªæµãã§æ¯ãè¿ãããšãã§ããŸãã
â ãã¬ãŒã³ãç»é¢é²ç»ããªãã宿œãã
Zoomã®é²ç»æ©èœããWindowsã®ç»é¢é²ç»ïŒWin+GïŒãªã©ã§ããã¬ãŒã³è³æãããããªãã話ããŠããæ§åãé²ç»ããŸãã
â¡ é²ç»ãã¡ã€ã«ãWhisperã§æåèµ·ãããã
SRT圢åŒã§åºåããããšã§ãèªåã話ããå
容ã«ã¿ã€ã ã¹ã¿ã³ããä»ããŸãã
⢠SRTãã¡ã€ã«ãAIïŒClaude ã ChatGPTïŒã«ã¢ããããŒããã
ããã®ãã¬ãŒã³ã®äžã§ã説æãäžååãªç®æãæ¹åãã¹ããã€ã³ããææããŠãã ããããšãé¡ããããšãAIãã¿ã€ã ã¹ã¿ã³ãä»ãã§ãã£ãŒãããã¯ããããŸãã
⣠ææãããå Žé¢ãFFmpegã§åãåºã
AIãã00:02:30ãããã®èª¬æãåããã«ããããšææããŠãããããFFmpegã䜿ã£ãŠãã®å Žé¢ã®ã¹ã¯ãªãŒã³ã·ã§ãããåç»ã¯ãªãããåãåºããŸãã
ffmpeg -ss 00:02:30 -i ãã¬ãŒã³é²ç».mp4 -frames:v 1 å Žé¢ãã£ããã£.pngãã®ã³ãã³ãäžè¡ã§ãæå®ããæéã®ã¹ã¯ãªãŒã³ã·ã§ãããåããŸãã
†å
容ãšããžã¥ã¢ã«ãã»ããã§æ¯ãè¿ã
ãäœã話ããŠãããïŒããã¹ãïŒããšããã®ãšãäœãæ ã£ãŠãããïŒã¹ã¯ãªãŒã³ã·ã§ããïŒããã»ããã§ç¢ºèªã§ããã®ã§ãæ¯ãè¿ãã®è³ªããŸã£ããå€ãããŸãã
ãã®æ¹æ³ã®ã¡ãªãã
ãã®ãSRT à FFmpeg à AIãã®çµã¿åããã䟿å©ãªçç±ã¯ãWhisperãå°å ¥ããæç¹ã§ãå¿ èŠãªããŒã«ããã¹ãŠæã£ãŠãããšããããšã§ãã
Whisper â SRTãã¡ã€ã«ïŒã¿ã€ã ã¹ã¿ã³ãä»ãããã¹ãïŒãçæ
FFmpeg â Whisperå°å ¥æã«ãã§ã«ã€ã³ã¹ããŒã«æžã¿ãåç»ããã®å Žé¢åãåºãã«äœ¿ãã
AIïŒClaude / ChatGPTïŒ â SRTãã¡ã€ã«ãèªã¿èŸŒãã§ãå 容ã®åæããã£ãŒãããã¯ãã§ãã
ã€ãŸããæ°ããããŒã«ã远å ããå¿ èŠããªããã§ãããã§ã«æå ã«ããéå ·ã®çµã¿åããã ãã§ããããŸã§ã§ãããšããã®ãå¬ãããã€ã³ãã§ãã

ãŸãšã
Whisperã¯ãç¡æã®æåèµ·ããããŒã«ããšããŠéåžžã«åªç§ã§ãããTXTãã¡ã€ã«ã ãã䜿ã£ãŠããŠã¯ãã£ãããªãã§ãã
ç¹ã« SRTãã¡ã€ã« ã«æ³šç®ãããšãã¿ã€ã ã¹ã¿ã³ãæ å ±ãæŽ»ãããŠããããªäœ¿ãæ¹ãã§ããŸãã
ãã¬ãŒã³ç·Žç¿ã®æ¯ãè¿ããšæ¹åç¹ã®ç¹å®
é²ç»ãã¡ã€ã«ããã®éèŠå Žé¢ã®èªååãåºã
AIãšçµã¿åããããå 容åæïŒããžã¥ã¢ã«ç¢ºèªã®ã¯ãŒã¯ãããŒ
ããããWhisperãã»ããã¢ããããæç¹ã§FFmpegãå ¥ã£ãŠããã®ã§ã远å ã³ã¹ããŒãã§å§ããããŸãã
ãã²ãã¿ãªããããTXT以å€ãã®åºå圢åŒã«ç®ãåããŠãWhisperã®æŽ»çšã®å¹ ãåºããŠã¿ãŠãã ããã