Master your delivery. Master your material. Master the room.
Elevare is an advanced, AI-powered multimodal public speaking coach that unifies emotion recognition, presentation analysis, dynamic question generation, answer validation, and feedback synthesis into a single, cohesive platform.
Nearly 50% of students report high levels of public speaking anxiety, and professionals routinely lose opportunities not due to a lack of expertise, but due to ineffective delivery. Existing tools tackle this problem in isolated pieces — pace tracking or filler-word detection or generic quiz questions. Elevare brings these together into one content-aware training experience that reacts to your material, your delivery, and your weaknesses.
Built as a graduation project at Ain Shams University — Faculty of Computer & Information Sciences.
| Capability | Description |
|---|---|
| 🗣️ Multimodal Emotion Recognition | Fuses speech, facial expression, and body language analysis into one unified delivery score |
| 📊 Objective + Empathetic Feedback | Dual-Agent architecture separates cold analytical scoring from warm, human-like coaching |
| 📑 Content-Aware Q&A | Uploads your slides and generates real, context-grounded practice questions — not generic templates |
| 🧠 Cognitive-Level Targeting | Questions rephrased across all six levels of Bloom's Taxonomy (Remember → Create) |
| ✅ Automated Answer Grading | Zero-shot LLM scoring with detailed, actionable feedback on every response |
| 👁️ Eye Contact & Gaze Tracking | Lightweight geometric gaze estimation, no heavy gaze-specific model required |
| Module | Approach | Result |
|---|---|---|
| Speech Emotion Recognition | Stacked Bi-LSTM over 190 acoustic features (MFCCs, Chroma, Log-Mel, etc.), PCA-reduced | 92.9% accuracy |
| Facial Emotion Recognition | GiMeFive CNN (5 conv blocks, ~10.5M params) on RAF-DB | 84.46% accuracy |
| Body Language Recognition | Temporal CNN over multi-scale joint/bone/velocity streams (MEED dataset) | 89.5% accuracy |
| Question Generation | Semantic Graph Conversational QG (SG-CQG) + Bloom's Taxonomy persona rephrasing via Llama 3.3-70B | High anchor grounding (~97–99%) |
| Answer Validation | Zero-shot ASAS-F-Z via DSPy Chain-of-Thought + Llama 3.3-70B | 72.48% grading accuracy |
| Feedback Generation | Dual-Agent (Gemini 2.5 Flash): Reasoning Agent + Coaching Agent, Sandwich + OIS method | Structured, bias-controlled reports |
Machine Learning / AI
- PyTorch, TensorFlow/Keras — deep learning models (BiLSTM, CNN, ConvNeXt, Temporal CNN)
- DSPy — declarative prompt orchestration for LLM evaluation
- Llama 3.3-70B (via Groq API), Gemini 2.5 Flash — LLM-driven generation & reasoning
- faster-whisper — speech-to-text transcription
- MediaPipe, OpenCV — facial landmark & gaze tracking
- Librosa — audio feature extraction
- Tesseract OCR — slide text extraction
Application
client/— frontend applicationserver/— backend API & AI pipeline orchestration
Main/
├── client/ # Frontend application
├── server/ # Backend services, AI pipeline, models
├── .gitattributes # Git LFS config for model weights
├── .gitignore
└── README.md
- Python 3.10+
- Node.js (for the client)
- Git LFS (large model files are tracked via LFS)
# Clone the repository
git clone https://github.com/GP-Elevare/Main.git
cd Elevare
# Pull LFS-tracked model files
git lfs pull
# Backend setup
cd server
npm install
# Frontend setup
cd ../client
npm install# Start the backend
cd server
node index.js
# Start the frontend
cd client
npm run devDeveloped by students of the Computer Science Department, Faculty of Computer & Information Sciences, Ain Shams University:
- Nouran Haitham Othman
- Malak Khaled Mohammed
- Malek Ahmed Mohamed
- Mohammed Wael Marwan
- Mohammed Akram Mohammed
- Abd-Rhman Osama Mohammed
Under the supervision of:
- Dr. Hanan Hindy — Lecturer, Computer Science Department
- TA. Radwa Reda Hossieny — Assistant Lecturer, Scientific Computing Department
The full graduation project documentation — covering related work, system design, implementation details, and experimental results — is available in docs/Documentation.pdf.
- 🎓 Real-world academic integration and pilot testing with university seminars
- 🌍 Multilingual support (starting with Arabic)
- 🧑🤝🧑 Immersive audience simulation via reactive virtual avatars