{ "cells": [ { "cell_type": "markdown", "id": "54c01a23", "metadata": {}, "source": [ "# πŸ‡°πŸ‡· Korean subs β€” build once, run forever\n", "\n", "**Every session:** `Runtime β†’ Change runtime type β†’ L4 GPU`, then `Runtime β†’ Run all`.\n", "\n", "- A box will pop up asking for YouTube links β€” paste them in. Or drop audio/video files into `MyDrive/korean_subs/inbox/` from your Mac.\n", "- Finished `.srt` files land in `MyDrive/korean_subs/subs/`. Processed media moves to `done/` (keep it for asbplayer).\n", "- The model (~3 GB) downloads to your Drive **once**. After that it just loads.\n", "- Already-finished files are skipped, so re-running never redoes work." ] }, { "cell_type": "code", "execution_count": null, "id": "ba19836d", "metadata": {}, "outputs": [], "source": [ "# 1 β€” Setup (Drive + packages; pip cache lives on Drive so reinstalls are quick)\n", "from google.colab import drive\n", "drive.mount('/content/drive')\n", "\n", "import os\n", "BASE = '/content/drive/MyDrive/korean_subs'\n", "for d in ['inbox', 'subs', 'done', 'models', 'pipcache']:\n", " os.makedirs(f'{BASE}/{d}', exist_ok=True)\n", "\n", "!pip install -q --cache-dir {BASE}/pipcache -U faster-whisper \"yt-dlp[default]\"\n", "\n", "# YouTube now needs a JavaScript runtime (deno) for yt-dlp\n", "import subprocess\n", "if not os.path.exists('/root/.deno/bin/deno'):\n", " subprocess.run('curl -fsSL https://deno.land/install.sh | sh -s -- -y', shell=True, capture_output=True)\n", "os.environ['PATH'] = '/root/.deno/bin:' + os.environ['PATH']\n", "\n", "COOKIES = f'{BASE}/cookies.txt'\n", "print('πŸͺ cookies.txt found' if os.path.exists(COOKIES) else 'πŸͺ no cookies.txt in korean_subs/ (YouTube may block Colab)')\n", "print('βœ… setup done')" ] }, { "cell_type": "code", "execution_count": null, "id": "56ead004", "metadata": {}, "outputs": [], "source": [ "# 2 β€” Settings + ask for links\n", "PREFER_HUMAN_SUBS = True # if the video already has real (non-auto) Korean subs, use those instead of Whisper\n", "DOWNLOAD_VIDEO = False # True = save the full video (for asbplayer); False = audio only (smaller, faster)\n", "\n", "MODEL = 'large-v3' # most accurate. Alternative to try if a file comes out badly: 'large-v2'\n", "BEAM_SIZE = 5 # higher = slightly more accurate, slower\n", "\n", "# A natural Korean sentence nudges Whisper toward proper spacing and punctuation.\n", "INITIAL_PROMPT = 'μ•ˆλ…•ν•˜μ„Έμš”. μ˜€λŠ˜μ€ 날씨가 정말 μ’‹λ„€μš”. 그럼 μ‹œμž‘ν•΄ λ³ΌκΉŒμš”?'\n", "\n", "print('Paste YouTube links one at a time and press Enter after each.')\n", "print('Press Enter on an empty box when you are done (or right away to only process inbox/).')\n", "URLS = []\n", "while True:\n", " link = input(f'Link {len(URLS) + 1}: ').strip()\n", " if not link:\n", " break\n", " URLS.append(link)\n", "print(f'βœ… {len(URLS)} link(s) queued')\n" ] }, { "cell_type": "code", "execution_count": null, "id": "2f002a58", "metadata": {}, "outputs": [], "source": [ "# 3 β€” Download from YouTube (human subs if available, otherwise media β†’ inbox/)\n", "import re, glob, yt_dlp\n", "\n", "def safe(name):\n", " return re.sub(r'[\\\\/:*?\"<>|]+', '_', name).strip()[:120]\n", "\n", "def already_done(stem):\n", " return bool(glob.glob(f'{BASE}/subs/{glob.escape(stem)}*.srt'))\n", "\n", "def with_cookies(opts):\n", " if os.path.exists(COOKIES):\n", " opts['cookiefile'] = COOKIES\n", " return opts\n", "\n", "def fetch(url):\n", " with yt_dlp.YoutubeDL(with_cookies({'quiet': True, 'skip_download': True})) as y:\n", " info = y.extract_info(url, download=False)\n", " stem = f\"{safe(info['title'])} [{info['id']}]\"\n", " if already_done(stem):\n", " print(f'⏭ already have subs: {stem}'); return\n", "\n", " human = [k for k in (info.get('subtitles') or {}) if k.startswith('ko')]\n", " if PREFER_HUMAN_SUBS and human:\n", " opts = {'quiet': True, 'skip_download': True,\n", " 'writesubtitles': True, 'writeautomaticsub': False,\n", " 'subtitleslangs': [human[0]],\n", " 'outtmpl': f'{BASE}/subs/{stem}.%(ext)s',\n", " 'postprocessors': [{'key': 'FFmpegSubtitlesConvertor', 'format': 'srt'}]}\n", " with yt_dlp.YoutubeDL(with_cookies(opts)) as y: y.download([url])\n", " print(f'πŸ“ human subs saved: {stem}')\n", " if not DOWNLOAD_VIDEO: return # subs done; no need for media\n", "\n", " fmt = 'bv*[height<=720]+ba/b[height<=720]/b' if DOWNLOAD_VIDEO else 'bestaudio/best'\n", " opts = {'quiet': True, 'format': fmt, 'outtmpl': f'{BASE}/inbox/{stem}.%(ext)s'}\n", " if DOWNLOAD_VIDEO: opts['merge_output_format'] = 'mp4'\n", " with yt_dlp.YoutubeDL(with_cookies(opts)) as y: y.download([url])\n", " print(f'⬇️ downloaded: {stem}')\n", "\n", "for url in URLS:\n", " try:\n", " fetch(url)\n", " except Exception as e:\n", " print(f'❌ {url}\\n {e}')\n", " print(' Blocked? Put cookies.txt in MyDrive/korean_subs/, or download on your Mac and drop the file into inbox/.')\n" ] }, { "cell_type": "code", "execution_count": null, "id": "7870e5ca", "metadata": {}, "outputs": [], "source": [ "# 4 β€” Transcribe everything in inbox/ with maximum-accuracy settings\n", "import shutil, time, ctranslate2\n", "from faster_whisper import WhisperModel\n", "\n", "MEDIA = ('.mp3', '.m4a', '.wav', '.mp4', '.mkv', '.webm', '.opus', '.flac', '.aac', '.ogg', '.mov')\n", "\n", "def ts(t):\n", " total = int(round(t * 1000))\n", " h, r = divmod(total, 3_600_000); m, r = divmod(r, 60_000); s, ms = divmod(r, 1000)\n", " return f'{h:02}:{m:02}:{s:02},{ms:03}'\n", "\n", "def split_segment(seg, max_dur=6.0, max_gap=0.8):\n", " # Break long Whisper segments into short subtitle lines using word timings\n", " words = seg.words or []\n", " if not words:\n", " yield seg.start, seg.end, seg.text.strip(); return\n", " cur = []\n", " for w in words:\n", " if cur:\n", " gap = w.start - cur[-1].end\n", " dur = w.end - cur[0].start\n", " sentence_end = cur[-1].word.strip()[-1:] in '.?!' and cur[-1].end - cur[0].start >= 1.5\n", " if gap > max_gap or dur > max_dur or sentence_end:\n", " yield cur[0].start, cur[-1].end, ''.join(x.word for x in cur).strip()\n", " cur = []\n", " cur.append(w)\n", " if cur:\n", " yield cur[0].start, cur[-1].end, ''.join(x.word for x in cur).strip()\n", "\n", "files = sorted(f for f in os.listdir(f'{BASE}/inbox') if f.lower().endswith(MEDIA))\n", "todo = [f for f in files if not already_done(os.path.splitext(f)[0])]\n", "print(f'{len(todo)} file(s) to transcribe')\n", "\n", "if todo:\n", " gpu = ctranslate2.get_cuda_device_count() > 0\n", " if not gpu:\n", " print('⚠️ No GPU! Runtime β†’ Change runtime type β†’ T4 GPU. (CPU works but is extremely slow.)')\n", " print('Loading model (first run downloads ~3 GB to Drive)…')\n", " model = WhisperModel(MODEL, device='cuda' if gpu else 'cpu',\n", " compute_type='float16' if gpu else 'int8',\n", " download_root=f'{BASE}/models')\n", "\n", " for f in todo:\n", " src, stem = f'{BASE}/inbox/{f}', os.path.splitext(f)[0]\n", " print(f'\\n🎧 {f}')\n", " t0 = time.time()\n", " segments, info = model.transcribe(\n", " src,\n", " language='ko',\n", " beam_size=BEAM_SIZE,\n", " best_of=5,\n", " temperature=[0.0, 0.2, 0.4, 0.6, 0.8, 1.0], # retries hard spots instead of guessing\n", " condition_on_previous_text=False, # prevents runaway repetition loops\n", " initial_prompt=INITIAL_PROMPT,\n", " vad_filter=True, # skips silence/music β†’ fewer hallucinated lines\n", " vad_parameters={'min_silence_duration_ms': 500},\n", " word_timestamps=True, # tighter subtitle timing\n", " )\n", " tmp = f'/content/{stem}.srt'\n", " with open(tmp, 'w', encoding='utf-8') as out:\n", " n = 0\n", " for seg in segments:\n", " for start, end, text in split_segment(seg):\n", " if not text: continue\n", " n += 1\n", " out.write(f'{n}\\n{ts(start)} --> {ts(end)}\\n{text}\\n\\n')\n", " print(f'\\r {seg.end / info.duration:5.1%}', end='')\n", " shutil.move(tmp, f'{BASE}/subs/{stem}.srt')\n", " shutil.move(src, f'{BASE}/done/{f}')\n", " print(f'\\r βœ… {n} lines in {(time.time() - t0) / 60:.1f} min β†’ subs/{stem}.srt')\n", "\n", "print('\\nπŸŽ‰ all done')\n" ] } ], "metadata": { "accelerator": "GPU", "colab": { "gpuType": "T4", "provenance": [] }, "kernelspec": { "display_name": "Python 3", "name": "python3" } }, "nbformat": 4, "nbformat_minor": 5 }