YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
SE-Probe model artefacts
Companion model checkpoints for SE-Probe, the public code release for "Where Does Speech Enhancement Adapt? Probing Study Under Controlled Degradation" (Amar, Ivry, Cohen, 2026).
π Paper: arXiv:2512.00482
Files
muse_reverb_e48.pt(2.3 MB): Reverb fine-tuned MUSE checkpoint (epoch 48). PESQ 3.02 / STOI 0.944 on the 824x5 RIR test set; this is the model probed in Section 4.2 of the paper. Trained from the upstream noise-only MUSE g_best, fine-tuned on RIR-Mega convolutive mixtures.
Dereverberation fine-tunes probed in the TASLP version (held-out test set: 824 utterances x 5 of 202 held-out RIR-Mega RIRs, 4,120 mixtures; reverberant input PESQ 1.658):
| file | model | test PESQ / STOI / SI-SDR (dB) |
|---|---|---|
cluster-2026-07/mpsenet_dereverb_e50/g_00295909 (+ config.json) |
MP-SENet, all parameters fine-tuned, epoch 50 | 3.361 / 0.966 / +9.44 |
cluster-2026-07/demucs_dereverb_v2_e57/g_00047310 |
Demucs DNS64, fine-tuned (Demucs upstream learning rate), epoch 57 | 2.441 / 0.918 / +3.77 |
cluster-2026-07/demucs_dereverb_e45/g_00037350 |
Demucs DNS64, fine-tuned (MUSE-recipe learning rate), epoch 45 | 2.339 / 0.914 / +3.34 |
cluster-2026-07/<arm>__seed<S>/g_00040000 |
MUSE profile-guided fine-tuning arms (full_ft, freeze_encoder, freeze_decoder; seeds 1234, 2345, 3456) |
see training/README.md |
Load them with se_probe.mpsenet.model.load_mpsenet_model(checkpoint_path=...) and se_probe.demucs.model.load_demucs_model(checkpoint_path=...). The training pipeline that produced these checkpoints is vendored in the SE-Probe repository under training/.
Citation
@article{amar2026seprobe,
title = {Where Does Speech Enhancement Adapt? Probing Study Under Controlled Degradation},
author = {Yair Amar and Amir Ivry and Israel Cohen},
year = {2026},
eprint = {2512.00482},
archivePrefix = {arXiv},
primaryClass = {eess.AS},
url = {https://arxiv.org/abs/2512.00482}
}
License
MIT.