Video Subtitle

Video Subtitle Pipeline / 视频字幕流水线

Complete workflow for adding subtitles to videos: audio extraction → speech-to-text → translation → subtitle burning

完整的视频字幕生成流程: 提取音频 → 语音转文字 → 翻译 → 烧录字幕

Read references/style-presets.md for the full ASS style reference. Read references/troubleshooting.md when something doesn't render correctly.

Prerequisites / 前置条件

• ffmpeg and ffprobe in PATH — verify with ffmpeg -version
• Python 3 with openai-whisper — pip install openai-whisper
• macOS / Linux / Windows
• First Whisper run downloads a model (~140MB for base)

Run npx video-subtitle-verify (or node bin/verify.js) at any time to check your environment.

快速开始 / Quick Start

Tell your agent: "add English subtitles to lecture.mp4" or "帮我把 lecture.mp4 生成中英双语字幕".

The agent will automatically invoke the pipeline. The four steps are described below in case you want to call them manually.

Workflow

Step 1 — Extract Audio / 提取音频

python3 scripts/extract_audio.py "<VIDEO_PATH>" --output "<AUDIO_PATH>.wav"

If --output is omitted, defaults to <video_stem>.wav. Output is 16kHz mono WAV — best for Whisper accuracy.

Step 2 — Transcribe (Whisper) / 语音转写

python3 scripts/transcribe.py "<AUDIO_PATH>" --model base --output "<SUBTITLE_PATH>.srt"

Models: tiny (fastest) · base (good default) · small · medium · large (best accuracy). First run downloads the model (~140MB for base); subsequent runs use the cache. The script prints the detected language — use this to decide if translation is needed.

Step 3 — Translate (AI-driven) / 翻译

If the detected language differs from the target, translate the SRT in-place:

1. Read the generated .srt to get all segments with timestamps.
2. Translate each segment's text to the target language, preserving the SRT structure exactly (index, timestamps unchanged).
3. Write the translated SRT to <original_stem>.<target_lang>.srt.

For multi-language or batch translation, dispatch to the subtitle-translator subagent (see .claude-plugin/agents/subtitle-translator.md).

Auto-open the editor after translating / 翻译后自动打开编辑器: Immediately after writing the translated SRT, launch the visual editor so the user can review and tweak the result without an extra step:

# With the source video (full pipeline):
python3 scripts/serve_preview.py "<VIDEO>" "<TRANSLATED>.srt"

# SRT-only (no video handy) — editor opens for text editing & Save SRT:
python3 scripts/serve_preview.py "<TRANSLATED>.srt"

This is the recommended hand-off: translation → editor → (optional) re-export. The editor auto-opens http://127.0.0.1:5173/ in the browser. If no video is passed, the editor runs in SRT-only mode (video preview hidden, Export/burn disabled) — the user can still edit text and download the corrected .srt.

Step 4 — Burn Subtitles / 烧录字幕

4a. Single Subtitle Mode / 单语字幕

python3 scripts/burn_subtitle.py "<VIDEO_PATH>" "<SUBTITLE_PATH>.srt" \
    --output "<OUTPUT>.mp4" --style default

Style presets: default (general) · streaming (Netflix-style) · white_shadow (cinematic) · yellow_black (high contrast) · minimal (least intrusive). See references/style-presets.md for the full ASS reference. Custom styles:

--custom "FontSize=20,PrimaryColour=&H00FFFFFF,OutlineColour=&H00000000"

4b. Dual Subtitle Mode (Original + Translation) / 双语字幕

python3 scripts/burn_dual_subtitle.py "<VIDEO_PATH>" "<ORIGINAL>.srt" "<TRANSLATED>.srt" \
    --output "<OUTPUT>.mp4"

Generates a combined ASS file (auto-cleaned) with two styles:

• Original (top): PingFang SC 18pt white, semi-transparent black background, MarginV=60
• Translated (bottom): Helvetica 16pt light green-white, semi-transparent black background, MarginV=18

Step 5 — Visual Editor (auto-opened after Step 3, optional) / 可视化编辑器

After translating (Step 3) the editor is launched automatically — this is the recommended hand-off. You can also open it any time to preview, tweak, and re-export the burned video — no manual ffmpeg flags needed. The editor runs entirely on a local server; nothing is uploaded.

# Lightweight web preview (recommended; works even without ffmpeg libass)
python3 scripts/serve_preview.py "<VIDEO>" "<SUBTITLE>.srt"

# SRT-only (e.g. right after translating, no video handy)
python3 scripts/serve_preview.py "<SUBTITLE>.srt"

# Optional: customize export filename / style / port
python3 scripts/serve_preview.py "<VIDEO>" "<SUBTITLE>.srt" \
    -o "<OUTPUT>.mp4" -s yellow_black --port 8080 --no-browser

• Auto-opens http://127.0.0.1:5173/ in your browser.
• Left: video player with a live bottom subtitle overlay.
• Right: every subtitle segment. Edit text, start/end timestamps, delete a segment, or click Seek to jump the video to that line.
• + Add segment inserts a blank line at the current playback time.
• Reload SRT re-reads the original file; Save SRT downloads the edited .srt; Export video burns the edited subtitles and downloads the MP4.
• Style presets: default, yellow_black, minimal.
• Export uses scripts/burn_moviepy.py (moviepy + Pillow) instead of ffmpeg's subtitles filter, so it works on ffmpeg builds without libass.

This step replaces Step 4 when you prefer visual editing over CLI burning.

FlyCut alternative (richer UI): scripts/serve_editor.py launches the FlyCut Caption editor (React SPA at editor/flycut/) in "bridge" mode. It supports dual-language editing (pass --translated "<TRANSLATED>.srt"), and routes export to the burn_* scripts. Options: --port <n>, --host <addr>, --no-browser. It requires building the SPA first (cd editor/flycut && pnpm install && pnpm build) and a ffmpeg build that supports libass, so the lightweight serve_preview.py is the safer default. Maintainer note: the editor page lives at editor/flycut/ — a copy of FlyCut Caption extended with a "bridge" mode (?bridge=1) that loads /video + /srt from this server and routes export to /api/burn. The built SPA is in editor/flycut/dist/. If you edit editor/flycut/src/, rebuild with cd editor/flycut && pnpm install && pnpm build.

Complete Pipeline Example / 完整示例

User says: "帮我把 lecture.mp4 生成中英双语字幕"

1. scripts/extract_audio.py  lecture.mp4                    → lecture.wav
2. scripts/transcribe.py     lecture.wav --model base       → lecture.srt
   (script prints: Detected language: zh)
3. AI translates lecture.srt → lecture.english.srt
   → automatically opens the editor for review:
     python3 scripts/serve_preview.py lecture.mp4 lecture.english.srt
4. scripts/burn_dual_subtitle.py lecture.mp4 \
       lecture.srt lecture.english.srt                      → lecture.dual.mp4

Platform Notes

• macOS: PingFang SC and Helvetica are universally available.
• Linux: Install fonts-noto-cjk (Debian/Ubuntu) or noto-fonts-cjk (Arch). Pass --custom "FontName=Noto Sans CJK SC,..." if needed.
• Windows: Prefer Microsoft YaHei or SimHei. Pass via --custom "FontName=Microsoft YaHei,...".
• Sandboxed paths (Desktop, Downloads): write intermediates to /tmp/ first, then copy the final output back. The shell script examples/full-pipeline.sh does this automatically.
• Whisper cache: models are cached in ~/.cache/whisper/ — only the first run downloads.

Notes

• ffmpeg's subtitles filter uses libass. ASS font names must match installed system fonts.
• BorderStyle=4 + BackColour creates a semi-transparent background box behind each subtitle line, improving readability on bright footage.
• For detailed troubleshooting, see references/troubleshooting.md.
• For style customization, see references/style-presets.md.

故障排查 / Troubleshooting

License / 许可证

MIT — see LICENSE.