LZC Camera crying detection assets

Immutable model and runtime assets for the experimental 8 kHz camera workflow. No camera recordings, user data, ESC-50 recordings, or OpenSLR recordings are stored here.

BEATs crying classifier (plugin 0.2.0)

beats-crying-2s.xml and .bin form one OpenVINO 2026.0.0 CPU IR model. It accepts a [1,198,128] normalized Kaldi fbank from a 2-second, mono 16 kHz waveform and returns the maximum AudioSet score for Whimper (81), Baby cry (395), and Crying,sobbing (498). The plugin accepts camera 8 kHz A-law and upsamples it to 16 kHz before fbank extraction; this does not restore frequencies above 4 kHz.

The source is Microsoft's BEATs iter3+ AS2M fine-tuned cpt2 checkpoint, SHA-256 e5815275a04b6885e7b8af63d120b29bffae2cd2225cf4915e1ec6d819d3022c. The transformer and classifier were traced from its fbank output and converted with openvino.convert_model. NNCF 2.19.0 mixed-preset post-training INT8 quantization used 96 two-second windows from 24 public clips, balanced across infant crying, adult crying, environmental sounds, and other human sounds. The app's fbank implementation was numerically compared with the checkpoint's TorchAudio preprocessing. The original full-waveform trace gave different scores and is not used. Source code and BEATs model are distributed under the accompanying Microsoft MIT license (BEATS-LICENSE.txt).

XML: SHA-256 9a43a295f524d25b80b3671bbf343466e0385b090004f1df3df182fa58858408, 1,453,311 bytes.
BIN: SHA-256 1b5903e3ce6157ea66ebe70769e3c08f453dce3a549ba05c153e7efec11cfc7e, 91,187,634 bytes.

The plugin analyzes 2-second windows every 0.48 seconds and by default confirms an event after 2 of the most recent 3 windows reach its standard score threshold of 0.24. This threshold was selected on the same exploratory data used for evaluation; it is not a household calibration.

Previous YAMNet asset

1.tflite is Google's YAMNet classification-tflite version 1, redistributed without modification under Apache 2.0. Original: https://www.kaggle.com/models/google/yamnet/tfLite/classification-tflite/1 . Implementation: https://github.com/tensorflow/models/tree/master/research/audioset/yamnet . SHA-256: 10c95ea3eb9a7bb4cb8bddf6feb023250381008177ac162ce169694d05c317de, 4,126,810 bytes. The new plugin does not load this file.

Runtime

runtime-cp312-x86_64.tar.gz contains unmodified wheel payloads for CPython 3.12 Linux x86_64: NumPy 2.2.6, OpenVINO 2026.0.0, and openvino-telemetry 2025.2.0. Component licenses remain in their .dist-info directories. The archive includes wheels.sha256 with original wheel hashes. SHA-256: ac6aa446111e800dd39833a9a29ab5a8a16c6cefe46fd2a81beafefd976072c0, 69,256,922 bytes. Sources: https://pypi.org/project/numpy/2.2.6/ , https://pypi.org/project/openvino/2026.0.0/ , https://pypi.org/project/openvino-telemetry/2025.2.0/ .

Downloads must be pinned to a repository commit and checked against the listed hashes before loading executable files. These artifacts alone do not establish household detection reliability or phone notification delivery.

Downloads last month: 26