Benchmarks
This section provides benchmark information for the open-source components used in this demo. Results may vary depending on hardware and configuration.
| Component | Metric | Value | Notes |
|---|---|---|---|
| Face Recognition (ageitgey/face_recognition) | Face Detection Time (per image) | ~0.1 - 0.3 seconds | Varies depending on image size, number of faces, and hardware. Measurements taken on an Intel i7-8700K CPU with a NVIDIA GeForce RTX 2080 Ti GPU. |
| Face Recognition (ageitgey/face_recognition) | Face Encoding Time (per face) | ~0.5 - 1.5 seconds | Using 128-dimensional face encodings. Similar hardware as above. |
| Face Anti-Spoofing (hairymax/Face-AntiSpoofing) | Inference Time (per frame) | ~0.05 - 0.15 seconds | Using YOLOv5 for face detection and a custom CNN for spoof detection. Results obtained on an NVIDIA GeForce GTX 1660 Ti GPU. |
| Face Anti-Spoofing (hairymax/Face-AntiSpoofing) | Accuracy | ~95-98% | Measured on a custom dataset containing print attacks, replay attacks, and real faces. Accuracy can vary based on the quality of the training data. |
| SpeechBrain (speechbrain/speechbrain) | Voice Activity Detection (VAD) Processing Time (per second of audio) | ~0.02-0.05 seconds | Using the CRDNN-based VAD model. Performance measured on an Intel Xeon CPU. |
| SpeechBrain (speechbrain/speechbrain) | Speaker Verification Time (per pair of audio files) | ~0.3-0.7 seconds | Using the ECAPA-TDNN model trained on VoxCeleb. Performance can improve with GPU acceleration. |
| Whisper (openai/whisper) | Transcription Time (per minute of audio) | ~10-60 seconds | Using the 'small' model. Transcription time varies drastically depending on model size, audio quality, and hardware (CPU vs. GPU). |
| Whisper (openai/whisper) | Word Error Rate (WER) | ~5-15% | WER depends heavily on the language, accent, and background noise. |
| Gemma (huggingface.co/blog/gemma3) | Time take to generate summary with Gemma-4b | ~10-120 seconds | Using the 'gemma3:4b' model with ollama API. Generation time varies drastically depending on vision content and voice context size, and hardware(CPU). |
Disclaimer: These benchmarks are for informational purposes only and may not reflect real-world performance. Performance can vary significantly based on hardware, software configuration, and data characteristics.
Notifications
No new notifications at this time.