Presentation Video Generator

The Problem It Solves

Creating a presentation video often involves synchronizing PDF slides with a script in Markdown format to produce a complete video with subtitles. Manual processes require juggling multiple tools—such as video editors, TTS services, and subtitle add-ons—which are cumbersome and error-prone. This is especially challenging when precise audio-visual alignment, handling long sentences, or enabling voice cloning is needed. For developers seeking automation, an integrated tool can minimize repetitive effort.

How It Works: Core Capabilities and Key Steps

The ppt-to-video skill provides an end-to-end pipeline. Key steps include:

• Step 0: Format the Script. Uses an LLM to convert user-provided scripts of any format into a standard Markdown structure, ensuring headers follow the ## Page X pattern for subsequent processing.
• Step 1: Extract Vocals (Optional). If reference audio is provided, uses Demucs to separate vocals and automatically selects the best segment (5-10 seconds) for voice cloning, supporting extraction from video sources.
• Step 2: Process the Script. Parses Markdown, cleans formatting (e.g., removes **bold** and *italic*), splits sentences (handling long ones), and generates TTS-friendly text, including vocalizing directory paths (e.g., ~/.codebuddy/skills/ becomes "home directory codebuddy skills directory").
• Step 3: Speech Synthesis. Supports two TTS engines: prioritizes Qwen3-TTS (local, based on Alibaba's Qwen, supporting voice cloning) or falls back to edge-tts (Microsoft's free online service with preset voices). The engine is auto-detected, generating per-sentence audio with pauses (default 0.4s between sentences, 1.0s between slides), and recording precise timing metadata in timing_data.json.
• Step 4: Generate Slide Video. Offers two modes: Traditional mode uses FFmpeg to create video from PDF screenshots; Remotion mode requires native TSX components to redesign visuals, ensuring animation frames align with timing_data.json timestamps to avoid low-resolution screenshots and animation limitations.
• Step 5: Merge Components. Combines video, audio, and subtitles into the final final_presentation.mp4. Subtitles default to white text with PingFang SC font and black stroke, supporting burned-in or soft subtitle modes.

The skill includes resume-from-breakpoints, process locks, and engine consistency checks for pipeline stability. For instance, re-running after interruption skips completed TTS audio.

Limitations and Considerations

• Input Requirements: PDF slides and a page-divided Markdown script (e.g., ## Page 1) are mandatory. The script must adhere to standard format; it's recommended to run an input validation script to check page count matching and formatting.
• Dependencies: Core dependencies include FFmpeg (for audio-video processing), Python 3.8+, and specific packages (e.g., PyMuPDF for PDF conversion). TTS engines require installing qwen-tts or edge-tts; optional tools include Whisper (for speech-to-text in Step 1) and Demucs (vocal separation).
• Performance and Resources: Qwen3-TTS local execution needs significant memory (8GB+ recommended), while edge-tts requires internet access. Video generation time depends on slide count and resolution (default 1920x1080).
• Audio-Video Sync: Ensure timing_data.json timing data is consistent, using 24kHz sample rate. Subtitle rendering relies on system fonts; Linux systems may need font substitution like Noto Sans CJK SC.
• Regeneration: Supports re-generating single TTS sentences via the --regenerate parameter, automatically updating timing data and re-merging the video.

This skill offers a flexible and powerful automation tool for technical users, but attention to environment configuration and dependency management is necessary to avoid common issues like TTS loading failures or missing fonts.