Bilibili Video Summary Tool#

Extract full content from a Bilibili video — transcript/subtitles, danmaku, comments, and description — then use your own LLM capabilities to produce a deep summary. No external AI API required (no OpenAI / Gemini key needed).

Capabilities#

Data SourceMethodPriorityCC SubtitlesBilibili APIFastest, used if availableAudio Transcriptionwhisper.cpp + Vulkan GPUAutomatic fallback when no subtitlesVideo Descriptionyt-dlpAlways capturedDanmaku (scrolling comments)yt-dlpParsed, analyzed for frequent contentCommentsBilibili Comment APIHot-sorted, deduplicated, top liked extracted

Workflow#

When you receive a Bilibili video link and are asked to summarize it, follow these steps:

Step 1: Extract all data#

bash

python bili-transcript.py "<video_url>"

The script automatically:

1. Gets video title, uploader, duration, description
2. Attempts Bilibili CC subtitles (fastest, used if available)
3. Falls back to GPU transcription: download audio → convert to wav → whisper.cpp with Vulkan
4. Downloads and analyzes danmaku (scrolling comments)
5. Fetches video comments, sorted by likes

Output files are saved to ./bili-output/:

• transcript.txt — full transcript/subtitle text
• danmaku.json — danmaku data with statistics
• comments.json — comment data with top-liked

The JSON output includes preview text, danmaku summary, and top comments.

Step 2: Read full transcript#

The JSON preview truncates at 2000 characters. Read the full file:

bash

cat ./bili-output/transcript.txt

Step 3: Read danmaku and comments#

Review community response data:

bash

cat ./bili-output/danmaku.json
cat ./bili-output/comments.json

Step 4: Compose your summary#

Use your own LLM capabilities to produce a comprehensive summary. Suggested structure:

Video Overview — Title, uploader, duration, transcription source (subtitle / GPU). Key info from the description (project links, update notes, etc.).

Core Content — What the video is about. Fluent paragraph summary of the main narrative.

Key Points — Notable arguments, data points, or information worth highlighting.

Community Response (optional) — Reactions from danmaku and comments. Skip if content is insubstantial (spam, trolling, no valuable discussion).

• Danmaku analysis: look for frequently repeated phrases (community memes/reactions), informative questions, technical discussions, controversy points
• Comment analysis: look for top-liked opinions, creator interactions, user-reported issues, technical insights

Assessment (optional) — Content quality, information density, notable strengths or weaknesses.

Available Actions#

bash

# Video metadata only
python bili-transcript.py "<URL>" --action info

# CC subtitles only (if available)
python bili-transcript.py "<URL>" --action subtitle

# Force GPU transcription (skip subtitle check)
python bili-transcript.py "<URL>" --action transcribe

# Danmaku only
python bili-transcript.py "<URL>" --action danmaku

# Comments only
python bili-transcript.py "<URL>" --action comments

# Custom output directory
python bili-transcript.py "<URL>" --output ./my-output

Environment Variables#

VariablePurposeWHISPER_CPP_DIRPath to whisper.cpp directory (containing whisper-cli)WHISPER_MODELPath to whisper model file (e.g., ggml-large-v3-turbo.bin)BILI_OUTPUT_DIRDefault output directory (default: ./bili-output)

Performance Reference#

Video LengthTotal TimeNotes5 minutes15sGPU transcription is fast12 minutes22sDownload + convert + transcribe1 hour2-3 minDepends on audio densityDanmaku/Comments5-10sDepends on comment volume

Dependencies#

• Python packages: yt-dlp, av (PyAV)
• Transcription engine: whisper.cpp with Vulkan support (optional, only needed if no CC subtitles)
• Model: ggml-large-v3-turbo.bin (~1.6GB, download separately)
• GPU: Any Vulkan-compatible GPU (NVIDIA, AMD, Intel) — auto-detected
• No external AI API keys required

Limitations#

• Requires internet access to Bilibili
• Some content requires login (paid courses, restricted videos) — may fail
• Danmaku and comment APIs may be rate-limited
• whisper.cpp does not support m4a; script auto-converts via PyAV
• Very long videos (>2 hours) take significant transcription time; try --action subtitle first
• Comments are fetched from the first 3 pages (~60 comments); may not cover very hot videos fully