AI-powered pipeline for generating YouTube Shorts videos (script → voice → image → final video)

EN:
This project builds a fully automated AI pipeline for short-form video content:

• script generation (OpenAI)
• thumbnail prompt + image generation
• voice generation via F5-TTS (local, free)
• smart chunking + audio merging
• final video rendering (MoviePy)

No paid TTS. No external voice services. Everything runs locally except text generation.

shorts-agent/

├── assets/
│ └── voice_clean.wav

├── output/
│ └── dev/
│ ├── script.json
│ ├── thumbnail_prompt.json
│ ├── thumbnail.png
│ ├── voice_chunk_1.wav
│ ├── voice_chunk_2.wav
│ ├── ...
│ ├── voiceover_en.wav
│ └── final_video_en.mp4

├── src/
│ ├── agents/
│ │ ├── script_agent.py
│ │ ├── image_prompt_agent.py
│ │ └── image_generator_agent.py
│ │
│ ├── prompts/
│ ├── schemas/
│ ├── utils/
│ │
│ ├── pipeline.py
│ ├── tts_generator.py
│ └── video_builder.py

git clone https://github.com/yourname/shorts-agent.git
cd shorts-agent

python -m venv .venv

Windows:
.venv\Scripts\activate

pip install -r requirements.txt

OPENAI_API_KEY=your_api_key_here

F5_MODEL=F5TTS_v1_Base
F5_REF_AUDIO=assets/voice_clean.wav
F5_REF_TEXT=Hello, this is a clean reference voice sample for text to speech generation.

• Fully local
• No cost per generation
• Uses reference audio + text
• Generates voice in chunks for stability

Reference audio must be:

• clean voice (no noise)
• single speaker
• 5–10 seconds
• calm tone (no эмоции, крики)
• must match F5_REF_TEXT

Script → Voice (chunked) → Merge → Image → Video

python src/pipeline.py -i

This generates:

• script
• thumbnail prompt
• thumbnail image
• voice chunks
• merged voice file

python src/video_builder.py -e

Output:

output/dev/final_video_en.mp4

Instead of generating one long audio:

• text is split into chunks
• each chunk → separate .wav
• then merged into one final audio

Example:

voice_chunk_1.wav
voice_chunk_2.wav
voice_chunk_3.wav
→ voiceover_en.wav

• ~120–180 characters per chunk
• prevents F5 crashes
• keeps voice natural

• MoviePy based
• caption overlay
• smooth zoom animation
• attaches merged audio
• exports final mp4

Check:

• voiceover_en.wav exists
• audio duration is not 0
• MoviePy uses with_audio()
• ffmpeg installed

Fix:

• reduce chunk size
• clean reference audio
• match reference text

Cause:

• bad reference audio

Fix:

• record clean voice sample

python src/pipeline.py

python src/pipeline.py -i
python src/video_builder.py -e

• Python
• OpenAI API
• F5-TTS (local)
• MoviePy
• FFmpeg
• Pillow
• python-dotenv

This is not just a generator.

It is a content factory:

• scalable
• cheap
• automatable
• ready for YouTube Shorts / TikTok