English | 简体中文 | 日本語 | Français | 한국어

Video Transcribe

Extract text/transcripts from YouTube and Bilibili videos. Supports CC subtitle extraction (fastest) and ASR transcription with LLM post-correction.

Features

• YouTube + Bilibili support: auto-detect platform from URL
• Subtitle-first strategy: tries CC subtitles before ASR (10-100x faster)
• Multiple ASR engines: faster-whisper (multilingual), Qwen3-ASR (Chinese-optimized)
• LLM post-correction: fixes ASR errors, technical terms, proper nouns
• Bilibili fallback: yt-dlp HTTP 412 -> automatic yutto fallback

Requirements

• Python 3.8+
• FFmpeg (must be in PATH)

Installation

# Core dependencies
pip install yt-dlp yutto faster-whisper torch openai

# Optional: Qwen3-ASR (better for Chinese)
pip install qwen-asr

# Optional: Bilibili login for CC subtitles and 1080P+
pip install playwright && playwright install

Usage

# Basic usage (auto-detect platform, try subtitles first)
python video_transcribe.py "https://www.youtube.com/watch?v=..."

# Bilibili video
python video_transcribe.py "https://www.bilibili.com/video/BV..."

# Specify output directory
python video_transcribe.py "URL" --output-dir ./transcripts

# With specific ASR model
python video_transcribe.py "URL" --asr-model whisper-small

# With Qwen3-ASR (better for Chinese)
python video_transcribe.py "URL" --asr-model qwen3-asr --qwen3-path /path/to/Qwen3-ASR-1.7B

# Skip LLM correction
python video_transcribe.py "URL" --no-llm-correction

# With custom LLM for correction
python video_transcribe.py "URL" --llm-model gpt-4o-mini --api-key YOUR_KEY

Parameters

Parameter	Default	Description
--output-dir	.	Output directory
--lang	zh	Language code for ASR
--asr-model	auto	ASR model (auto/whisper-small/whisper-base/whisper-medium/whisper-large/qwen3-asr)
--whisper-size	small	Whisper model size
--qwen3-path		Path to local Qwen3-ASR model
--no-llm-correction	false	Skip LLM post-correction
--llm-model	gpt-4o-mini	LLM model for correction
--api-key		API key (or set OPENAI_API_KEY env var)
--base-url		Base URL for OpenAI-compatible API

How It Works

URL -> detect platform
  |-- YouTube -> yt-dlp
  +-- Bilibili -> yt-dlp (fallback: yutto on HTTP 412)
       |-- CC subtitles available? -> extract (fastest)
       +-- No subtitles -> download audio -> ASR
            |-- Chinese + Qwen3 available -> Qwen3-ASR (recommended)
            +-- Other / no Qwen3 -> faster-whisper
                 +-- LLM correction -> output markdown

Output Format

# Video Title

**Source**: URL
**Method**: CC Subtitle / Qwen3-ASR / faster-whisper/small
**Date**: 2025-01-01 12:00
**Corrected**: Yes/No
**Characters**: 12345

---

Full corrected transcript text here...

Bilibili Login (for CC subtitles and 1080P+)

CC subtitles and high-quality audio often require login:

# Login via Playwright (opens browser for QR code scan)
python -c "
import asyncio
from playwright.async_api import async_playwright

async def login():
    async with async_playwright() as p:
        browser = await p.chromium.launch(headless=False)
        page = await browser.new_page()
        await page.goto('https://passport.bilibili.com/login')
        print('Scan QR code or log in...')
        await page.wait_for_url('https://www.bilibili.com/**', timeout=120000)
        cookies = await page.context.cookies()
        for c in cookies:
            if c['name'] in ('SESSDATA', 'bili_jct', 'DedeUserID'):
                print(f\"{c['name']}={c['value']}\")
        await browser.close()

asyncio.run(login())
"

# Pass cookies to yt-dlp
yt-dlp --cookies-from-browser chrome "URL" ...

License

MIT