The music-understanding brain of Egystormy Quickie, a free, open-source track organizer for DJs. This model powers the Find Similar Tracks engine: give it a track, it returns a 768-dimension embedding, and tracks whose embeddings are close sound close.
Converted by the Egystormy project from m-a-p/MERT-v1-95M (Apache-2.0) with a direct PyTorch → ONNX export (opset 14). Faithfulness was verified numerically against the original PyTorch model (pooled-embedding cosine ≈ 1.000000 on the test signal).
input_values — mono waveform resampled to 24 kHz, float32,
shape [batch, samples]. Quickie uses a 10-second segment.last_hidden_state, shape [batch, time, 768].File: mert-v1-95m.onnx (fp32, ~361 MB), sha256
0e9a681bdd63938b3718827d85a48595a1f014b9dfcce6d7ef93237316ba2a32.
Part of the Egystormy open-source project — free for everyone.