Creating whiteboard-style educational videos demands precise coordination between freehand illustrations and spoken narration, yet no existing method addresses this multimodal synchronization problem with structured, reproducible drawing representations. We present the first dataset of 24 paired Excalidraw demonstrations with narrated audio, where every drawing element carries millisecond-precision creation timestamps spanning 8 STEM domains. Using this data, we study whether a vision-language model (Qwen2-VL-7B), fine-tuned via LoRA, can predict full stroke sequences synchronized to speech from only 24 demonstrations. Our topic-stratified five-fold evaluation reveals that timestamp conditioning significantly improves temporal alignment over ablated baselines, while the model generalizes across unseen STEM topics. We discuss transferability to real classroom settings and release our dataset and code to support future research in automated educational content generation.
Figures & tables
Statistic
Value
Demonstrations
24
Domains
8
Total elements
877
Elements per demo (mean / min / max)
36.5 / 16 / 79
Total drawing span (s)
814.2
Per-demo span (mean s)
33.9
Table 1: ExcaliTeach dataset statistics computed from dataset_stats.json and aligned transcripts.
Setting
TAE ↓
Chamfer ↓
Type Acc. ↑
E1 (Full)
0
0
0
E2 (No onset token)
0
0
0
E3 (No canvas image)
0
0
0
E4 (One-shot)
0
0
0
E5 (No transcript)
0
0
0
Table 2: Automatic evaluation across E1–E5 (5-fold mean). Lower is better for TAE and Chamfer; higher is better for Type Acc.
Setting
Gemini Judge ↑
Human Preference (%) ↑
E1 (Full)
0
0
E2 (No onset token)
0
0
E3 (No canvas image)
0
0
E4 (One-shot)
0
0
E5 (No transcript)
0
0
Table 3: Holistic quality evaluation across E1–E5.