mirth/earshot-clap-htsat-unfused

Earshot CLAP bundle laion-htsat-unfused-2d293ecc875b

ONNX export of laion/clap-htsat-unfused (revision 8fa0f1c6d0433df6e97c127f64b2a1d6c0dcda8a) for on-device sound detection in Earshot.

• audio_encoder.onnx: audio encoder, FP32, input log-mel [1, 1, 1001, 64].
• text_encoder.onnx: text encoder, int8 weights (onnxruntime MatMulNBits + int8 embeddings).
• preprocessing.json, tokenizer.json, model.json: everything needed to reproduce the inputs.
• manifest.json: file sizes and SHA-256 checksums; verify every file against it before use.

Weights are unmodified apart from export and the text-encoder quantization. The original model is by LAION and is licensed Apache-2.0; this export is distributed under the same license.

File	Role	Size	SHA-256
audio_encoder.onnx	audio_encoder	117.3 MB	4ef210b195dc43daebcdb1b19897c8b7a48acbac9c16dbbd7fb82d481d9b06ab
model.json	model_info	0.0 MB	1a9f8915b2d4312cf893d08ef3c2de2078b4e8308e919c4aac5a97ad2d90717c
preprocessing.json	preprocessing	0.0 MB	f609186a2507e2896c9b7feaa76c7ba26adc073aa56e3c34f246b414df334da6
text_encoder.onnx	text_encoder	140.2 MB	71021fbb1f1bc8d29a0d5b0ce69c76808d0e063a60daa7dfcd7bf5142cec7b46
tokenizer.json	tokenizer	3.6 MB	712ec8f3a640c1d1aca341eb7d6b13df128e6dd735ae020ccf879719cf5701cb