lecture-mind 0.2.0

Event-aware lecture summarizer using V-JEPA

Lecture Mind

Event-aware lecture summarizer using V-JEPA visual encoder for real-time, context-aware summaries and retrieval.

Features

• Visual Encoding: DINOv2 ViT-L/16 for 768-dim frame embeddings
• Text Encoding: sentence-transformers (all-MiniLM-L6-v2) for query embeddings
• Audio Transcription: Whisper integration for lecture transcription
• Multimodal Search: Combined visual + transcript ranking with configurable weights
• Event Detection: Automatic slide transition and scene change detection
• FAISS Index: Fast similarity search with IVF optimization for large collections

Installation

Basic Installation (CPU)

pip install lecture-mind

With ML Dependencies (GPU recommended)

pip install lecture-mind[ml]

With Audio Transcription

pip install lecture-mind[audio]

Full Installation

pip install lecture-mind[all]

Development Installation

git clone https://github.com/matte1782/lecture-mind.git
cd lecture-mind
pip install -e ".[dev,ml,audio]"

Quick Start

CLI Usage

# Process a lecture video
lecture-mind process lecture.mp4 --output data/

# Query the processed lecture
lecture-mind query data/ "What is gradient descent?"

# List detected events
lecture-mind events data/

# Get help
lecture-mind --help

Python API

from vl_jepa import (
    VideoInput,
    FrameSampler,
    VisualEncoder,
    TextEncoder,
    MultimodalIndex,
    EventDetector,
)

# Load and sample video frames
with VideoInput.from_file("lecture.mp4") as video:
    sampler = FrameSampler(fps=1.0)
    frames = sampler.sample(video)

# Encode frames (uses placeholder encoder by default)
encoder = VisualEncoder.load()
embeddings = encoder.encode_batch(frames)

# Build searchable index
index = MultimodalIndex()
index.add_visual(embeddings, timestamps=[f.timestamp for f in frames])

# Query the lecture
text_encoder = TextEncoder.load()
query_emb = text_encoder.encode("machine learning basics")
results = index.search(query_emb, k=5)

for result in results:
    print(f"Timestamp: {result.timestamp:.1f}s, Score: {result.score:.3f}")

Architecture

lecture.mp4
    │
    ▼
┌─────────────┐     ┌─────────────┐     ┌─────────────┐
│ VideoInput  │────▶│FrameSampler│────▶│   Frames    │
└─────────────┘     └─────────────┘     └─────────────┘
                                               │
                    ┌──────────────────────────┼──────────────────────────┐
                    ▼                          ▼                          ▼
            ┌─────────────┐            ┌─────────────┐            ┌─────────────┐
            │VisualEncoder│            │EventDetector│            │AudioExtract │
            │  (DINOv2)   │            │             │            │  (FFmpeg)   │
            └─────────────┘            └─────────────┘            └─────────────┘
                    │                          │                          │
                    ▼                          ▼                          ▼
            ┌─────────────┐            ┌─────────────┐            ┌─────────────┐
            │  Embeddings │            │   Events    │            │ Transcriber │
            │  (768-dim)  │            │             │            │  (Whisper)  │
            └─────────────┘            └─────────────┘            └─────────────┘
                    │                          │                          │
                    └──────────────────────────┼──────────────────────────┘
                                               ▼
                                    ┌─────────────────┐
                                    │ MultimodalIndex │
                                    │     (FAISS)     │
                                    └─────────────────┘
                                               │
                                               ▼
                                    ┌─────────────────┐
                                    │  Search/Query   │
                                    └─────────────────┘

Performance

Operation	Target	Actual
Query latency (1k vectors)	<100ms	30.6µs
Search latency (100k vectors)	<100ms	106.4µs
Frame embedding (placeholder)	<50ms	0.36ms
Event detection	<10ms	0.24ms

See BENCHMARKS.md for detailed performance analysis.

Requirements

• Python 3.10+
• NumPy >= 1.24.0
• OpenCV >= 4.8.0

Optional Dependencies

• ML: PyTorch >= 2.0, transformers, sentence-transformers, FAISS
• Audio: faster-whisper >= 1.0.0
• UI (v0.3.0): Gradio >= 4.0.0

Development

# Run tests
pytest tests/ -v

# Run with coverage
pytest tests/ --cov=src/vl_jepa --cov-report=term

# Lint and format
ruff check src/ && ruff format src/

# Type check
mypy src/ --strict

# Run benchmarks
pytest tests/benchmarks/ -v --benchmark-only

Roadmap

• v0.1.0: Foundation (placeholder encoders, basic pipeline)
• v0.2.0: Real Models + Audio (DINOv2, Whisper, multimodal search)
• v0.3.0: User Experience (Gradio web UI, Docker)
• v1.0.0: Production (optimization, real decoder, deployment)

License

MIT License - see LICENSE for details.

Citation

If you use Lecture Mind in your research, please cite:

@software{lecture_mind,
  title = {Lecture Mind: Event-aware Lecture Summarizer},
  author = {Matteo Panzeri},
  year = {2026},
  url = {https://github.com/matte1782/lecture-mind}
}