AI视频短视频创作器

Two-phase AI-narrated video short creator with human review checkpoints. Phase 1 analyzes source materials, generates TTS narration via edge-tts, and exports a subtitle review sheet for user approval. Phase 2 assembles the final video with FFmpeg including clipping, scaling, xfade transitions, SRT subtitle burning, and audio merge. Use when creating a short-form video with AI voiceover from existing video clips and a script, especially for tech/AI paper explanation videos for social media.

Video Short Creator

Two-phase workflow for creating AI-narrated short videos from source clips and a script, with human review at every critical step.

Prerequisites

Tool	Purpose	Install
FFmpeg	Video processing	winget install FFmpeg or download from ffmpeg.org
Python 3.8+	Script execution	Use the managed runtime
edge-tts	Microsoft TTS	pip install edge-tts (no API key needed)
SimHei font	Chinese subtitle rendering	Usually pre-installed on Windows (C:\Windows\Fonts\simhei.ttf)

Phase 1: Material Preparation & Review

This phase generates the narration audio and exports a review sheet. Stop and wait for user approval before proceeding.

Step 1.1: Analyze Materials

1. Ask the user for:
   • Source video clips (paths or a folder)
   • Script / narration text (or generate one from provided reference material)
   • Target resolution (default: 1920x1080 landscape)
   • Voice preference (default: zh-CN-YunxiNeural, young male Mandarin)
   • Publishing platform (optional, affects aspect ratio recommendations)
2. Scan all source clips with ffprobe to catalog:
   • Resolution, duration, file size
   • Generate a clip inventory table
3. If no script is provided but reference material (paper URL, markdown, etc.) is given:
   • Research the content and draft a narration script
   • Structure as 3-5 segments with natural flow

Step 1.2: Build SCRIPT Data Structure

Create a Python list of segment dicts that maps each narration segment to source clips:

python

SCRIPT = [
    {
        "id": "seg1",          # unique segment ID
        "text": "...",         # narration text for this segment
        "videos": [            # source clips to use
            {"file": "clip1.mp4", "start": 0, "max_dur": 14},
            {"file": "clip2.mp4", "start": 5, "max_dur": 30},
        ],
    },
    # ... more segments
]

Key design decisions:

• Each segment's total available video duration should exceed narration duration by ~0.5s
• Use multiple clips per segment for visual variety
• max_dur prevents over-long clips from a single source

Step 1.3: Generate TTS & Export Review Sheet

Run scripts/step1_generate_review.py with the SCRIPT config:

bash

python scripts/step1_generate_review.py

This script:

1. Generates TTS audio for each segment using edge-tts
2. Extracts clean subtitle timing from SentenceBoundary events (NOT WordBoundary)
3. Exports subtitle_review.md with all subtitle entries + timing info

Step 1.4: Present Review Materials to User

Display the following for user review:

1. Clip Inventory - All source clips with metadata
2. Script Review - Full narration text organized by segment
3. Subtitle Review Sheet (subtitle_review.md) - All subtitle entries with:
   • Sequence number, segment ID, start/end time, duration, text content
4. Segment-to-Clip Mapping - Which clips are used for each narration segment

Ask the user to review and confirm or provide modifications:

• Text corrections in narration or subtitles
• Timing adjustments
• Clip substitutions or reorderings
• Adding/removing subtitle entries

CRITICAL: Do NOT proceed to Phase 2 until the user explicitly approves.

Phase 2: Video Assembly

Only start after user approval of Phase 1 materials.

Step 2.1: Apply User Edits (if any)

If the user provided subtitle edits:

1. Parse their feedback and update the subtitle entries accordingly
2. If they created a subtitle_edited.txt, load it (format: start_sec|end_sec|text per line)
3. Re-verify that all source clip paths exist

Step 2.2: Execute Video Assembly

Run scripts/step2_edit_video.py:

bash

python scripts/step2_edit_video.py

The assembly pipeline:

1. Clip & Scale - Extract clips at specified start/duration, scale to target resolution (1920x1080)
2. Concatenate - Join multiple clips per segment using concat demuxer
3. Burn Subtitles - Overlay subtitles using .srt file + FFmpeg subtitles filter
4. xfade Transitions - Chain segments with alternating fade/fadeblack transitions (0.8s)
5. Build Audio - Concatenate narration segments with short silences between them
6. Final Merge - Combine video + audio into final output

Step 2.3: Output

• Final video at output/{PROJECT_NAME}_FINAL.mp4
• Resolution: 1920x1080, 30 FPS
• Subtitle style: SimHei font, size 14, white text, thin outline, bottom-aligned (movie-style)
• Audio: AAC 128kbps
• Video: H.264 CRF 20

Step 2.4: Present Final Video

Use open_result_view to present the final video. Ask user for feedback on:

• Subtitle readability and positioning
• Transition smoothness
• Audio-video sync
• Overall pacing

Configuration

All configuration lives in the SCRIPT data structure at the top of each script. Key constants:

Constant	Default	Description
VOICE	zh-CN-YunxiNeural	edge-tts voice (Mandarin, young male)
TARGET_W/H	1920x1080	Output resolution
XFADE_DUR	0.8	Transition duration in seconds
SUBTITLE_FONT_SIZE	14	Subtitle font size
SUBTITLE_MARGIN_V	50	Bottom margin for subtitles

Known Pitfalls (Windows)

• Never use drawtext filter — shell escaping for \n and special characters is unreliable on Windows. Always use .srt file + subtitles filter instead.
• SRT file paths need escaping for FFmpeg subtitles filter: replace \ with / and : with \:.
• edge-tts SentenceBoundary gives clean text; WordBoundary may contain artifacts like stray n characters.
• Always re-fetch subtitle timing even when audio files are cached — edge-tts does not store boundary events.
• xfade filter labels must use tmp0, tmp1 format (not v01, v02) to avoid FFmpeg parsing errors.
• For detailed pitfalls and debugging guidance, see references/ffmpeg-pitfalls.md.