sireme/PPT_to_mp4_with_sample_voiceover

🎙️ PowerPoint Narrator --- AI Voice Cloning

Convert any .pptx presentation into a fully narrated MP4 video using your own voice, powered by Coqui XTTS v2.

🔗 Live Hugging Face Space:\ https://huggingface.co/spaces/sireme/PPTtomp4withsample_voiceover

✨ What It Does

Upload a PowerPoint file and a short voice sample, and the app automatically:

1. 1.Extracts text from each slide
2. 2.Converts PPTX → PDF → PNG (robust multi-slide rendering)
3. 3.Clones your voice using Coqui XTTS v2
4. 4.Generates natural narration for every slide
5. 5.Syncs slide duration to narration length
6. 6.Exports a clean H.264/AAC MP4 video at 24fps

No manual recording. No editing. Fully automated.

🧠 How It Works

Full Processing Pipeline

.pptx file ──┐ ├──▶ python-pptx ──▶ Slide Text ─────────────┐ │ │ └──▶ LibreOffice ──▶ PDF ──▶ PNG Slides ─────┤ ├──▶ moviepy ──▶ Final MP4 Voice sample ──▶ ffmpeg ──▶ Clean WAV ──▶ XTTS v2 ──▶ Narration WAVs ─────┘

Architecture Decisions

📄 Slide Rendering (Docker-Safe)

Direct PPTX→PNG export can fail in headless environments.\ This app uses:

PPTX → PDF → PNG (via poppler-utils)

This guarantees: - All slides captured correctly - No "1/1 slide" issue - Stable rendering inside Docker & HF Spaces

📝 Text Extraction

python-pptx extracts text from every shape and text frame on each slide.

• —All visible text is included
• —Long slides are chunked intelligently
• —Slides with no text remain silent (no forced placeholder narration)

🎤 Voice Processing

Voice samples are normalized using ffmpeg to:

• —Mono
• —22050 Hz
• —16-bit PCM WAV

This ensures compatibility and stable XTTS inference.

🗣 Voice Cloning (XTTS v2)

Model:\ tts_models/multilingual/multi-dataset/xtts_v2

Optimizations:

• —Model loaded once globally at startup
• —Speaker embedding computed once per request
• —Text chunked at sentence boundaries
• —Inference wrapped in torch.inference_mode()
• —Audio normalized and faded between chunks
• —GPU memory cleared after inference (if applicable)

🎬 Video Assembly

moviepy:

• —24 FPS output
• —H.264 video
• —AAC audio
• —Subtle Ken Burns zoom effect
• —Slide duration = narration duration

📦 Supported Inputs

Input Details ----------------------------------- ----------------------------------- Presentation .pptx files, up to 30 slides

Voice sample .wav, .mp3, .aac, .m4a, .flac, etc.

Languages en, es, fr, de, it, pt, pl, tr, ru, nl, cs, ar, zh-cn, hu, ko, ja -----------------------------------------------------------------------

Voice sample guidelines:

• —3--30 seconds
• —Clear speech
• —Minimal background noise
• —Natural pacing

🗂 Project Structure

├── app.py ├── Dockerfile ├── requirements.txt └── README.md

⚙️ Running Locally

Prerequisites

• —Docker installed
• —Optional: NVIDIA GPU for faster inference

Build

docker build -t pptx-narrator .

Run (CPU)

docker run -p 7860:7860 pptx-narrator

Run (GPU)

docker run --gpus all -p 7860:7860 pptx-narrator

Open in browser:

http://localhost:7860

🚀 Deploying to Hugging Face Spaces

1. 1.Create a new Docker Space
2. 2.Select Docker as SDK
3. 3.(Recommended) Enable T4 GPU
4. 4.Push:
5. 5.app.py
6. 6.Dockerfile
7. 7.requirements.txt
8. 8.README.md

On first boot, XTTS (\~1.8GB) downloads automatically.

The Docker container includes stability configurations:

• —Hugging Face cache redirected to /tmp
• —Numba caching disabled
• —Matplotlib config redirected
• —LibreOffice given writable HOME
• —Librosa patched to remove deprecated pkg_resources dependency

This prevents:

• —Slide export failures
• —Numba locator errors
• —pkg_resources import crashes
• —HF cache permission warnings

🏎 Performance

Hardware \~Time per 10 Slides -------------------- ---------------------- T4 GPU (HF Spaces) \~1--2 minutes CPU only \~8--15 minutes

Processing time scales roughly linearly with slide count.

The 30-slide cap prevents excessive latency.

📦 Dependencies

Python

• —TTS (Coqui XTTS v2)
• —torch / torchaudio
• —transformers
• —python-pptx
• —moviepy
• —gradio
• —soundfile

System

• —ffmpeg
• —libreoffice
• —poppler-utils

⚠️ Limitations

• —Slides with no text remain silent
• —Very dense slides may produce slower narration
• —English performs best
• —30 slide maximum per run

🔐 Responsible Use

Voice cloning must only be used with explicit consent.

Do not use this tool to:

• —Clone someone's voice without permission
• —Create deceptive content
• —Impersonate individuals for fraud

Intended for:

• —Personal presentation narration
• —Accessibility tooling
• —Productivity automation

📜 License

Uses Coqui XTTS v2 under the Coqui Public Model License (CPML).\ All other repository code is released under the MIT License.