Baby Cry Detector
Experimental quality-filtered compact CNN v0.2 for binary cry versus non_cry classification. The ONNX file is used by the local Next.js demo.
Audio contract
- mono, 16 kHz
- 49,152 samples / 3.072 seconds per window
- symmetric zero-padding for short clips
- 1.536-second hop for long clips, including a final end-aligned window
- pre-model application gate rejects near-silent or broadband-noise-like input
- RMS normalization target: -23 dBFS, limited to 12 dB gain and -1 dBFS peak
- log-mel frontend parameters and training statistics are stored in
model-metadata.json
Evaluation
| Evaluation set | Files | Accuracy | Balanced accuracy | Macro-F1 | Cry recall |
|---|---|---|---|---|---|
| Development validation | 207 | 99.52% | 98.96% | 99.32% | — |
| Regression partition | 204 | 97.06% | 95.19% | 95.85% | — |
| Enes cry-only probe | 294 | — | — | — | 97.28% |
The Enes probe contains only cry-positive examples, so its recall is not full classification accuracy. These are dataset-level evaluation results, not independent real-world validation.
Use and limitations
This model is for experimentation only. It must not be used for medical diagnosis, safety monitoring, or deciding whether an infant needs care. The score describes one processed clip and is not the model's overall accuracy.
License
The model artifact and repository-authored documentation are released under Creative Commons Attribution-ShareAlike 4.0 (CC BY-SA 4.0). No training audio is included. This license does not relicense third-party datasets; downstream users remain responsible for the original dataset terms and unresolved lineage.