Models, licences and credits
The Guardian OS demo runs the same console as the on-site system, on a 3D model of the PUP Main Building, Sta. Mesa. All detection happens inside your browser: audio is analysed in memory and never recorded or uploaded. These are the models it downloads, where they come from, and everything else it is built on.
On-device models
| Model | Source | Licence | In the demo |
|---|---|---|---|
| YAMNet sound classifier (float32) | Google, via MediaPipe model storage; runs on MediaPipe Tasks Audio 1.1.0 | Apache 2.0 | Gunshot, scream, shout and breaking-glass sounds every 0.5 s (4 MB) |
| Filipino FastConformer CTC, INT8 | NCSpeech/stt_tl_fastconformer_hybrid_large, ONNX export by OpenVoiceOS, CTC INT8 package by AladdinChen; used unchanged | CC BY 4.0 | Filipino and English keyword recognition on 3-second windows ("voices", 175 MB) |
| Pose Landmarker (lite) | Google, via MediaPipe model storage; runs on MediaPipe Tasks Vision 1.1.0 | Apache 2.0 | Body pose on the camera preview of Room S204, only while that preview is open; mapped to the hub's fall and crowd-low rules (6 MB) |
| Laya distress judge, fine-tuned | convaiinnovations/laya (multilingual; mmBERT-base encoder), fine-tuned by the Guardian team on its own template sentences, exported to ONNX with 8-bit weights | Apache 2.0 | Decides whether recognised speech is a real call for help ("AI judge", downloaded after the voices) |
The on-site system also uses Whisper as a second opinion, YOLO11 pose on CCTV feeds and weapon models. Those are not part of this browser demo; pose here uses MediaPipe on your own camera preview, with the hub's thresholds unchanged. Laya is fast but over-eager: on held-out tests it caught 15 of 16 handwritten threats and wrongly flagged 53 of 390 calm sentences, so a person confirms every alert before a lockdown.
Demo media
- 3D campus: "Polytechnic University of the Philippines" by Nico on SketchUp 3D Warehouse, exported from SketchUp and optimised with glTF-Transform; 3D Warehouse terms of use.
- Hallway CCTV loops: excerpts from the WiseNET (CC0) and CAVIAR (CC BY-SA; our excerpts CC BY-SA 4.0) research datasets, shown only as labelled backgrounds. Full footage credits.
- Demo clips: gunshot and everyday sounds from OpenGameArt, BigSoundBank and Freesound, all CC0.
- Alert chime: "Bell Chime Alert" by plasterbrain on Freesound, CC0 1.0.
- Brand: the Guardian OS Safe Passage logo, mark and icon were generated for this project.
Software
- onnxruntime-web 1.30.0 (MIT), MediaPipe Tasks Audio and Vision 1.1.0 (Apache 2.0), @huggingface/tokenizers 0.2.0 (Apache 2.0)
- React 18.3.1 and ReactDOM (MIT), htm 3.1.1 (Apache 2.0), three.js r160 (MIT)
- Geist and Geist Mono fonts (SIL Open Font License 1.1)
Privacy
Microphone audio stays in your browser's memory for a few seconds and is then discarded. Camera frames from the room preview are analysed for body pose one at a time and dropped; only joint positions reach the page. Speech transcripts live only on the open incident in this tab. Incident history is kept in this browser's local storage, without audio, transcripts or summaries, and expires after 90 days. The site uses no analytics or cookies. Model files are fetched once from models.reymar.app and cached by your browser.