youtube-transcriber

分类: 内容与多媒体 | 上传者: lifesized | 下载: 0 | 版本: v1.0（最新）

使用本地转录服务从 YouTube 视频捕获转录。该服务会在可用时获取官方 YouTube 字幕，或使用 Whisper 在本地进行转录。

更新日志: Source: GitHub https://github.com/lifesized/youtube-transcriber

目录结构

当前层级: tree/main/contrib/claude-code/

📄 README.md 1.3 KB
📄 SKILL-lite.md 8.2 KB
📄 SKILL.md 4.9 KB

SKILL.md

# YouTube Transcriber

Capture transcripts from YouTube videos using a local transcription service. The service fetches official YouTube captions when available, or transcribes locally using Whisper.

## When to Use

Use this skill when the user:
- Pastes a YouTube URL (with or without a verb like "transcribe", "summarize", "watch")
- Says "transcribe", "summarize", or "caption" followed by a YouTube URL
- Uses a shorthand like "s", "t", or "ts" followed by a YouTube URL (e.g. "ts https://youtube.com/watch?v=..." to transcribe and summarize)
- Asks to extract, capture, or get what was said in a YouTube video
- Sends just a YouTube URL with no other context

## Prerequisites

The YouTube Transcriber service must be running at `http://127.0.0.1:19720`.

To check if it's running:
```bash
curl -s http://127.0.0.1:19720/api/transcripts | head -c 100
```

If not running, the user needs to start it in their youtube-transcriber project directory:
```bash
npm run dev
```

## API Endpoints

Base URL: `http://127.0.0.1:19720`

### Capture a Transcript

```bash
curl -X POST 'http://127.0.0.1:19720/api/transcripts' \
  -H 'Content-Type: application/json' \
  -d '{"url": "YOUTUBE_URL"}'
```

**Response:**
```json
{
  "id": "abc123",
  "videoId": "dQw4w9WgXcQ",
  "title": "Video Title",
  "author": "Channel Name",
  "videoUrl": "https://youtube.com/watch?v=dQw4w9WgXcQ",
  "transcript": "[{\"text\":\"Hello\",\"startMs\":0,\"durationMs\":1500}]",
  "source": "youtube_captions",
  "createdAt": "2025-01-15T10:30:00.000Z"
}
```

The `transcript` field is a JSON string. Parse it to get an array of segments:
- `text`: The spoken words
- `startMs`: Start time in milliseconds
- `durationMs`: Duration in milliseconds

The `source` field indicates:
- `youtube_captions`: Fetched from YouTube (fast)
- `whisper_local`: Transcribed with Whisper (slower, 1-5 min)

### List All Transcripts

```bash
curl 'http://127.0.0.1:19720/api/transcripts'
```

### Search Transcripts

```bash
curl 'http://127.0.0.1:19720/api/transcripts?q=SEARCH_TERM'
```

### Get Single Transcript

```bash
curl 'http://127.0.0.1:19720/api/transcripts/ID'
```

### Delete Transcript

```bash
curl -X DELETE 'http://127.0.0.1:19720/api/transcripts/ID'
```

## Default Behavior

When the user uses "ts" (transcribe + summarize) or "s" (summarize), **always present a summary to the user**. When the user uses "t" (transcribe), return the full transcript unless they say otherwise. This is the expected workflow:

1. Capture the transcript via the API (saves it to the library)
2. Parse the transcript segments into plain text
3. Summarize the content directly — focus on key topics, insights, and takeaways
4. Present the summary with the video title and author

If the user asks for the full transcript, format it with timestamps instead of summarizing.

## Example Workflow

When user asks: "Transcribe this https://youtube.com/watch?v=abc123"

```bash
# 1. Capture the transcript
RESPONSE=$(curl -s -X POST 'http://127.0.0.1:19720/api/transcripts' \
  -H 'Content-Type: application/json' \
  -d '{"url": "https://youtube.com/watch?v=abc123"}')

# 2. Check for errors
if echo "$RESPONSE" | grep -q '"error"'; then
  echo "Error: $(echo "$RESPONSE" | jq -r '.error')"
  exit 1
fi

# 3. Extract metadata and transcript text
echo "$RESPONSE" | python3 -c "
import json, sys
resp = json.load(sys.stdin)
print(f'Title: {resp[\"title\"]}')
print(f'Author: {resp[\"author\"]}')
print(f'Source: {resp.get(\"source\", \"unknown\")}')
print()
segments = json.loads(resp['transcript'])
for seg in segments:
    print(seg['text'], end=' ')
print()
"
```

Then summarize the transcript text and present the summary to the user.

## Formatting Full Transcripts

When the user asks for the full transcript, format timestamps as `[MM:SS]`:

```bash
echo "$RESPONSE" | jq -r '.transcript' | python3 -c "
import json, sys
for seg in json.load(sys.stdin):
    ms = seg['startMs']
    print(f'[{ms//60000}:{(ms%60000)//1000:02d}] {seg[\"text\"]}')
"
```

## Error Handling

| Status | Meaning | Action |
|--------|---------|--------|
| 200/201 | Success | Process response |
| 400 | Invalid URL | Check URL format |
| 404 | Video not found | Verify URL is correct |
| 429 | Rate limited | Wait 1-2 minutes, retry |
| 500 | Server error | Check if service is running |
| Connection refused | Service not running | User needs to run `npm run dev` |

## Supported URL Formats

- `https://www.youtube.com/watch?v=VIDEO_ID`
- `https://youtu.be/VIDEO_ID`
- `https://www.youtube.com/shorts/VIDEO_ID`
- `https://www.youtube.com/embed/VIDEO_ID`
- URLs with additional parameters (playlists, timestamps)

## Notes

- Submitting a duplicate URL returns the existing transcript (no re-processing)
- Videos without captions use local Whisper transcription (can take several minutes)
- Transcripts are stored locally in SQLite and persist across sessions
- Always summarize by default — the full transcript is saved in the library for later

登录后下载/点赞/收藏 ❤ 8 | ★ 0

youtube-transcriber

目录结构

SKILL.md

举报内容

提示