CED: Consistent ensemble distillation for audio tagging
Paper • 2308.11957 • Published • 1
CED-Tiny with an embedding output (VisionState)
The ONNX file of mispeech/ced-tiny (revision
ace276d29dd0bb3f3517b0fa8cf300738c409019) with one output added, embedding: the encoder's
tokens pooled (mean, standard deviation, maximum; N × 576). The weights and the class scores
(logits, N × 527 AudioSet classes, already probabilities) are the original's. Input:
waveform, N × samples of 16 kHz mono audio.
Made by tools/soundmodel/make_model.py in https://github.com/oleost/VisionState for its sound
sensors, which recognise sounds such as sirens and smoke alarms and learn a household's own sounds
from the embedding.
CED: Xian Li, Nian Shao, Xiaofei Li, "CED: Consistent ensemble distillation for audio tagging", ICASSP 2024 (https://arxiv.org/abs/2308.11957). Licence: Apache-2.0, as the original. Trained by its authors on AudioSet.
Model tree for oleost/visionstate-sound
Base model
mispeech/ced-tiny