Skip to content

RysUpAlign learned alignment model v6

Latest

Choose a tag to compare

@rysupaudio-lab rysupaudio-lab released this 08 Oct 07:12

RysUpAlign learned alignment model v6: frame embeddings for audio-to-audio vocal alignment (HuBERT-base layers 1–9, frozen, + a 1.66 M-parameter head with a log-mel branch; 128-d, 100 fps). Also on Hugging Face: https://huggingface.co/rysupaudio/rysupalign-learned-v6

file sha256
rysupalign_learned_v6.onnx (fp32, 305 MB) db3bf5b88834bba926c147d8785e829845c7827248efc74e2966f0bec7ec8b32
rysupalign_learned_v6.int8.onnx (dynamic int8, 112 MB) e48d692579af3d8ed710ff99e47a1c37d0c1e0205440d5a69d1521a69f505246

Input wav [1, S]: mono 16 kHz, zero mean / unit variance. Output emb [1, T, 128], unit norm, 100 fps, frame k centred at 10k + 12.5 ms. Use with runners/learned_onnx.py or runners/m1_dtw.py (set RYSUPALIGN_ONNX or put the file in models/).

Results on RAB and their caveats: see the README.

Licence: the weights are under the PolyForm Noncommercial License 1.0.0 (MODEL_LICENSE.md); the HuBERT-base encoder weights they contain remain under the Apache License 2.0 (NOTICE, LICENSES/Apache-2.0.txt). Commercial licensing: https://rysupaudio.com/pages/contact-us

Required Notice: Copyright 2026 Rys Up Audio LLC (https://rysupaudio.com)