Commit History
4 days ago
-
Card from docs/models/ear.md
Verified signaturehttps://github.com/Desert-Ant-Labs/desert-ant-core/commit/e3bab7722187921dc1d279f1bd81a680ac28e47dc4936b6 -
Card from docs/models/ear.md
Verified signaturehttps://github.com/Desert-Ant-Labs/desert-ant-core/commit/5ecb1162b870b75f7178de2394e0336c6fe60e0d8550d92
7 days ago
-
Sync card from manifest.json
Verified signaturef13e0a5
8 days ago
-
Sync card from manifest.json
Verified signature5d385b0
15 days ago
-
Sync card from manifest.json
Verified signature9435686 -
Card: the blocks the manifest sync maintains
Verified signatureaca9062
17 days ago
-
Weights v0.2.0 on main, with the model card and third-party notices
Verified signature16aeb02
18 days ago
-
Sync card from manifest.json
Verified signatured0538a6
-
Sync card from manifest.json
Verified signaturefbb78b5
-
Sync card from manifest.json
Verified signature77d6f5c
-
Sync card from manifest.json
Verified signature4f1b3e1
-
Sync card from manifest.json
Verified signature25289b6
-
Sync card from manifest.json
Verified signaturea487dc6
-
Sync card from manifest.json
Verified signature239fca2
-
Sync card from manifest.json
Verified signature557fcde
-
Write the card: files, inputs and outputs, what it hears, accuracy, limits, languages, built on
Verified signature6202d3a -
Sync card from manifest.json
Verified signature364f7a3 -
Say audio in the tagline; Tongue is the text one
Verified signatureadf5bd9 -
Restore the card after the sync flattened it again
Verified signature014a22b -
Sync card from manifest.json
Verified signature54acf20 -
Generate the citation block; every card now has one
Verified signaturee89d35f -
Card pass: reader-first copy, generated install and license blocks
Verified signature67747e6 -
Restore the model card, and mark the generated header
Verified signaturef648f0c
-
Sync card from manifest.json
Verified signature60d9e79
-
Move the frontend into the SDK; one artifact per runtime
Verified signatureThe log-mel frontend is now host code rather than part of the graph. It cannot run in float16 -- the power spectrum is squared magnitudes floored at 1e-10 and 80% of its bins fall below float16's smallest normal number -- so shipping it as a Core ML program meant either destroying the features (routing 97.5% to 84.2%) or a float32 program that cannot use the Neural Engine at all. Both runtimes now take log-mel and return logits, so every platform feeds the same tensor to the same graph. mel_filters.f32 carries the filterbank so host code does not reimplement slaney-normalized librosa filters and drift from it silently. ear.mlmodelc: 256 ops, 100% Neural Engine, palette6-group32. ear.tflite: int8 with Hadamard rotation, signature input 'mel'.5ac5bec -
ear 0.1.0: spoken language identification, Core ML and LiteRT
Verified signatureDerived from openai/whisper-tiny (MIT), stripped to the subgraph that predicts a language: 37.8M parameters to 17.9M, identical predictions. Core ML ships as two artifacts on purpose. The frontend runs in fp32 because its power spectrum is squared magnitudes clamped at 1e-10 and 80% of those bins fall below fp16's smallest normal number; fusing it into the fp16 detector scores 84.2% routing where the split scores 97.5%. LiteRT is one file because it runs the frontend in fp32 anyway. Detector is 256 ops, 100% Neural Engine, 15.8 ms per 30 s window.47e2d59 -
initial commit
Verified signatured1827b8