From your recording we produce an annotated video plus charts so you can trace eye contact, body posture, speech rate and more.
VideoAnalytics is a research tool developed by the Centre for University Teaching at the University of Bayreuth. It enables lecturers and students to have video recordings of presentations or simulated exam dialogues analysed automatically.
The analysis includes body language, gestures, eye contact, speech behaviour and optionally AI-based emotion recognition. The goal is to receive constructive feedback for improving your own presentation skills.
Record your webcam and optionally your screen, then upload for analysis.
No software installation required. The compatibility check verifies in advance whether camera, microphone and browser audio are working.
Start recordingSubmit an existing recording (MP4, WebM, MOV) for analysis.
Upload an existing video file. The analysis starts automatically and you will receive your feedback by email.
Upload videoRecord your presentation via webcam or upload an existing video. Optionally, you can also record your screen at the same time.
Automatic evaluation of posture, gestures, gaze direction, speech pauses, volume and optionally facial expressions/emotions. Processing takes a few minutes.
You will receive an email with a link to your results including interactive charts, a transcribed text version and optionally an AI coaching report.
After the analysis you receive an interactive feedback dashboard. Here are examples of the different analysis sections:
Skeleton overlay shows posture, gestures and gaze direction in real time
Interactive charts with zoom, video cursor and calibration
Evidence-based assessment following scientific standards
Colour-coded improvement suggestions directly in the transcript
Detailed metrics with benchmark comparison
Personalised coaching feedback from Claude (Anthropic)
Combined assessment from voice, gaze, speech, gesture and facial expression — with moment highlights for confident passages and improvement areas
Video with text overlays at key moments — shows when confidence rises or drops
Label 10 diverse frames so the software accurately detects your individual gaze direction
Set start and end time so only the relevant section is analysed — no distorted data from walking in
Per analysis section: thumbs up/down + comments. Your feedback helps us improve the software
The optional emotion analysis uses a neural network (HSEmotion) to detect emotional states such as joy, surprise, concentration or tension from facial expressions. This data is presented exclusively in statistically aggregated form and serves for self-reflection.
Important: Emotion recognition is an approximation method. It does not capture inner emotional states but interprets visible facial expressions. The results should be understood as guidance, not as a psychological diagnosis.
Regulation (EU) 2024/1689 (AI Act) classifies emotion recognition systems in the workplace and educational institutions as particularly sensitive (Art. 5 para. 1 lit. f). Their use is only permitted under strict conditions.
The following safeguards apply within this project:
The use of this service is entirely voluntary. There is no obligation to record or upload videos for analysis. All optional features (emotion analysis, AI coaching) must be actively enabled.
All uploaded videos and analysis results are automatically deleted after 14 days. After that, neither the video nor the analysis data will be accessible. Early deletion can be requested at any time via the admin panel.
Processing takes place on a dedicated server at the University of Bayreuth. No data is shared with third parties. The AI coaching feature (optional) uses the Claude API from Anthropic; only the transcribed text is transmitted, no video recordings.
This website does not use cookies, tracking or external analytics services. No personal data is collected beyond the email address required for the upload.
VideoAnalytics combines specialised AI models and open-source libraries for comprehensive analysis:
| Category | Technology | Purpose |
|---|---|---|
| Body Pose & Gesture | MediaPipe (Google) | Body pose (skeleton landmarks), face mesh (468 points), iris tracking |
| Speech Recognition | faster-whisper (Systran) | Speech-to-text (transcript), word-based filler detection |
| Emotion Recognition | HSEmotion (HSE) | Facial emotion classification (joy, neutral, sadness, anger, etc.) |
| Gaze Direction | L2CS-Net (ResNet-34) | Gaze angle estimation (yaw = horizontal rotation, pitch = vertical tilt) from the face crop |
| Facial Expression (Action Units) | py-feat (Cosanlab) | Action Units = smallest visible facial-muscle movements defined by the Facial Action Coding System (FACS); py-feat detects them automatically |
| Voice Analysis | openSMILE (audEERING) | Acoustic features following the eGeMAPS standard (extended Geneva Minimalistic Acoustic Parameter Set): pitch, loudness, jitter = pitch fluctuation, shimmer = loudness fluctuation, HNR = Harmonics-to-Noise Ratio (ratio of harmonic content to noise) |
| Voice Quality | Parselmouth/Praat | Voice quality metrics based on Praat (the de-facto standard tool of phonetics research) |
| Filler Words (audio-based) | Eigenes Verfahren (ZHL UBT, auf eGeMAPS-Basis) | Custom method: detects “uh” / “um” from acoustic features (eGeMAPS) instead of from the transcript — more reliable than text-only detection |
| Emphasis & Three-Channel Coherence | Eigenes Verfahren (ZHL UBT) | Custom method: measures whether vocal emphasis, gesture and pause align while speaking (three-channel coherence) |
| Orientation Toward Presentation | Eigenes Verfahren (ZHL UBT) | Custom method: infers from hand, gaze and body direction when the speaker is facing the audience vs. the presentation (slides/board) |
| AI Coaching | Claude (Anthropic) | Speech quality analysis, personalised coaching — under a Zero-Data-Retention agreement: content is not stored and not used for AI training |
| Video Processing | OpenCV + FFmpeg | Per-frame processing, video encoding |
| Charts | Chart.js | Interactive timeline visualisation |
| GPU Acceleration (graphics-card computing) | NVIDIA CUDA 12.1 | AI-inference acceleration on NVIDIA graphics cards |
| Framework | FastAPI + Python | Backend interface (API = Application Programming Interface) |
University of Bayreuth
Centre for University Teaching (ZHL)
Universitätsstraße 30
95447 Bayreuth
Centre for University Teaching (ZHL)
University of Bayreuth
The University of Bayreuth is a public corporation. It is legally represented by the President.
Bavarian State Ministry of Science and the Arts
