Commit History
18 days ago
-
Weights v0.2.0 on main, with the model card and third-party notices
Verified signature16aeb02
-
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
