Video Subtitle Extractor 🎬→📝

Cross-platform multi-engine ASR subtitle extraction pipeline. Downloads audio from any yt-dlp-compatible video platform, transcribes with SenseVoice / whisper.cpp / openai-whisper, and applies LLM-based text calibration for Chinese content.

Default engine: SenseVoice Small (Alibaba FunASR) — ~1.5GB RAM, 234MB disk, 20× realtime speed, ~96% Chinese accuracy.

Tested & verified on Windows 11 with real Bilibili & Xiaohongshu videos.

Quick Start

# One-command full pipeline (SenseVoice Small — default, blazing fast for Chinese)
python scripts/run.py <video_url> --output-dir ./output

# Use whisper.cpp GGML (even lighter, 2GB RAM)
python scripts/run.py <video_url> --backend whispercpp --model medium-q5_1

# Use openai-whisper (standard, 5GB RAM)
python scripts/run.py <video_url> --backend openai --model medium

# Download audio only
python scripts/download_audio.py <video_url> <output_dir>

# Download audio + video (keep both as middleware)
python scripts/download_audio.py <video_url> <output_dir> --save-video --video-quality 1080

# Transcribe existing audio with auto backend selection
python scripts/transcribe.py <audio_file> --backend auto --language zh

# Transcribe with specific backend
python scripts/transcribe.py <audio_file> --backend sensevoice --language zh
python scripts/transcribe.py <audio_file> --backend whispercpp --model medium-q5_1
python scripts/transcribe.py <audio_file> --backend openai --model medium

When to Use This Skill

Use this skill when:

1. The video has no built-in subtitles (Bilibili, Xiaohongshu, YouTube, etc.)
2. You need high-accuracy Chinese transcription (~95% with medium model)
3. You want multiple output formats (TXT, SRT, VTT, JSON)
4. You need LLM-assisted text calibration for financial/technical terms
5. The user says: "下载字幕", "提取字幕", "语音转文字", "视频转文字", "字幕提取", "ASR转写"

Workflow

Step 0: Install Dependencies (once)

python scripts/install_deps.py

Auto-detects OS and installs: ffmpeg (winget/brew/apt), yt-dlp (pip), openai-whisper (pip). Handles Windows ffmpeg path detection even when not in PATH.

Step 1: Download Audio

Run scripts/download_audio.py <url> [output_dir].

Uses yt-dlp to extract the best available audio format (m4a preferred). Supports Bilibili, YouTube, and 1800+ yt-dlp-compatible platforms. The script automatically detects ffmpeg even when not in system PATH.

Optional: Download video as middleware

Add --save-video to persist the full video alongside audio:

python scripts/download_audio.py <url> --save-video --video-quality 1080
python scripts/run.py <url> --save-video --video-quality 720 --calibrate

--video-quality preset table:

⚠️ B站 note: Without login cookies, B站 caps at 480p. 720p+ requires --cookies-from-browser.

If download fails: the video may require cookies. Try:

yt-dlp --cookies-from-browser chrome <url>

Step 2: ASR Transcription (Multi-Backend)

Run scripts/transcribe.py <audio> --backend <engine> --model <size> --language <lang>.

Three backends, auto-selected by default (priority: SenseVoice → whisper.cpp → openai-whisper):

🥇 SenseVoice Small (default for Chinese)

Why SenseVoice? Alibaba's FunASR engine, Chinese-optimized from the ground up. No HuggingFace download needed (uses ModelScope mirror in China). ~4× faster than openai-whisper medium on CPU, with comparable or better Chinese accuracy.

🥈 whisper.cpp (GGML quantized, CPU-optimized)

Setup: pip install pywhispercpp + download GGML model from huggingface.co/ggerganov/whisper.cpp → place in ~/.cache/whispercpp/. See --show-backends for instructions.

🥉 openai-whisper (standard)

Note: small model removed (88-90% accuracy, superseded by SenseVoice). Use SenseVoice for light weight.

Step 3: Rule-Based Calibration

After transcription, apply calibrate.py for mechanical corrections:

# Calibrate a raw .txt transcript
python scripts/calibrate.py <output_dir>/<video_title>.txt
# Output: <video_title>_calibrated.txt

# Skip traditional→simplified conversion (already simplified input)
python scripts/calibrate.py raw.txt --no-tradsimp

Calibrate categories: homophone fixes, financial terms, AI/tech company names, semiconductor/hardware terms, traditional→simplified (600+ chars). For context-aware fixes (semantic errors, ambiguous names), use LLM review on top of rule-based output. See references/calibration_guide.md for the full 80+ pattern library.

Step 4: Deliver Results

Present the calibrated text. Always include:

• Model used (medium/large) and quality notes
• Any sections with low confidence or unclear audio
• Summary of corrections applied (counts by category)

Intermediate Artifacts & Step-by-Step Reuse

Every pipeline stage saves its output to disk. All artifacts persist in output_dir/ after the run — no data is lost between stages.

Artifact Map

Selective Reuse via run.py Flags

# Full pipeline (download + transcribe + calibrate)
python scripts/run.py <url> --calibrate --output-dir ./out

# Save video as middleware (downloads .m4a + .mp4)
python scripts/run.py <url> --save-video --calibrate --output-dir ./out

# Audio already exists → skip download, re-transcribe
python scripts/run.py <url> --skip-download --output-dir ./out

# Audio + transcript already exist → skip both, re-run calibration only
python scripts/run.py <url> --skip-download --skip-transcribe --calibrate --output-dir ./out

run.py auto-detects existing artifacts by matching the audio filename base. It will warn and fall back to downloading/transcribing if no match is found.

Standalone Scripts

Each stage has an independent entry point:

# Stage 1: Download audio only
python scripts/download_audio.py <url> [output_dir] [filename]

# Stage 1: Download audio + video at 720p (standalone)
python scripts/download_audio.py <url> [output_dir] --save-video --video-quality 720

# Stage 2: Transcribe existing audio
python scripts/transcribe.py <audio.m4a> --model medium --language zh --output-dir ./out

# Stage 3: Calibrate raw transcript (rule-based only)
python scripts/calibrate.py <raw.txt> [--output <path>] [--no-tradsimp]

Typical Reuse Scenarios

Scenario A — Change model, keep audio

# Already have .m4a from previous run
python scripts/run.py <url> --skip-download --model large-v3 --output-dir ./out

Scenario B — Change language, keep audio + transcript

# Have both .m4a and .txt; just recalibrate
python scripts/run.py <url> --skip-download --skip-transcribe --calibrate --language en --output-dir ./out

Scenario C — Batch calibrate multiple transcripts

# Apply calibration to all raw .txt files in a directory
Get-ChildItem .\out\*.txt | Where-Object { $_.Name -notmatch '_calibrated' } | ForEach-Object {
    python scripts/calibrate.py $_.FullName
}

Platform Support

Extending with New ASR Models

scripts/transcribe.py is designed for backend extensibility:

1. Add model info to ALL_MODELS dict
2. Implement transcribe_<backend>() function
3. Add CLI flag in argparse
4. Add backend to detect_backends()

Available backends: sensevoice (✅ production), whispercpp (✅ code ready, GGML model manual), openai (✅ production)

Backend auto-selection: When --backend auto (default), the engine picks the best available backend in priority order:

1. SenseVoice — Chinese-optimized, fastest, lightest
2. whisper.cpp — CPU-optimized, quantized models
3. openai-whisper — general purpose, most compatible

Troubleshooting

Performance Benchmarks (Tested)

Tested on Windows 11, Intel i7, 16GB RAM. Performance may vary by CPU speed.

Changelog

v2.0.0 — Multi-Engine ASR

• New: SenseVoice Small backend (Alibaba FunASR) — default engine, ~1.5GB RAM, 20× realtime, ~96% Chinese accuracy
• New: whisper.cpp GGML backend via pywhispercpp — CPU-optimized quantized models (0.5-2GB RAM)
• New: --backend parameter (auto/sensevoice/whispercpp/openai)
• New: transcribe.py --show-backends diagnostic command
• New: Auto-backend detection and fallback chain (SenseVoice → whisper.cpp → openai-whisper)
• Remove: small model from openai-whisper (superseded by SenseVoice)
• Remove: faster-whisper stub (CTranslate2 — WDAC incompatibility, never actually worked)
• Improve: transcriber architecture — clean backend dispatch, shared output writer
• Improve: Chinese-optimized default path (SenseVoice Small via ModelScope, no VPN needed)

v1.0.8

• New: Semiconductor / hardware domain calibration rules (30+ patterns) — 韬定律→道定律, 全站协同→全栈协同, 量子碎穿→量子隧穿, 吸片→芯片, 奈米→纳米, etc.
• Improve: Huawei 道定律 video now correctly calibrated (0→26 pattern corrections, ~98% accuracy)
• Fix: Douyin/TikTok platform note — clearly states cookies requirement

v1.0.7

• New: --video-quality parameter — presets best/2160/1440/1080/720/480/360 + raw yt-dlp format support
• New: find_artifacts_in_dir() replaces find_audio_in_dir() — caches both .m4a AND .mp4 on --skip-download --save-video
• Change: Video download format selector now uses height<= filters (graceful fallback) instead of hardcoded mp4-only
• Improve: download_video() now reports which quality preset is in use
• Improve: Artifact map includes video (.mp4) with quality column
• Improve: SKILL.md adds --video-quality usage table and B站 login caveats

v1.0.6

• New: --save-video flag — download and persist full video (.mp4) alongside audio
• New: download_video() function in download_audio.py (standalone: --save-video)
• Improve: Artifact map now includes video (.mp4) as first-class middleware
• Improve: _pipeline_meta.json includes video_path for full traceability

v1.0.5

• Remove: small model from all backends (88-90% accuracy, too poor for production)
• Change: Default model locked to medium (was medium in code, but docs still promoted small)
• Improve: Model table and benchmarks now medium-only baseline

v1.0.2

• New: Xiaohongshu (小红书) platform support — yt-dlp XiaoHongShu extractor
• New: Short link auto-resolution (xhslink.com → full URL via redirect)
• Improve: Platform support table now lists 小红书 explicitly
• Improve: Quick Start examples include xhslink.com usage

v1.0.1

• Fix: Expanded ffmpeg search paths from 3→7 (winget/scoop/chocolatey/ProgramFiles(x86))
• Fix: ensure_deps() now injects ffmpeg into os.environ['PATH'] on success
• Fix: SIGKILL troubleshooting updated — root cause is ffmpeg PATH, not OOM
• Improve: Auto-detect GPU (torch.cuda.is_available()) for fp16 support
• Improve: verbose=True for real-time transcription progress visibility
• Improve: More accurate error messages in dependency checks

v1.0.0

• Initial release: download (yt-dlp) + transcribe (whisper) + calibrate (LLM) pipeline
• 7 ffmpeg install path auto-detection
• Multi-format output (TXT, SRT, VTT, JSON)
• Platform support: Bilibili, YouTube, all yt-dlp sites