BS-Roformer-HyperACE

v2_inst--------------------------------------------------------------------<br> https://mvsep.com/quality_checker/entry/9475<br> <code> Metric sdr for instrum: 17.4007<br> Metric si_sdr for instrum: 17.3154<br> Metric l1_freq for instrum: 39.9287<br> Metric log_wmse for instrum: 14.1031<br> Metric aura_stft for instrum: 16.5670<br> Metric aura_mrstft for instrum: 22.8199<br> Metric bleedless for instrum: 37.8745<br> Metric fullness for instrum: 38.0255<br> </code> Currently, this model holds the highest aura_mrstft score on the instrumental side of the Multisong dataset.<br> <br> <br> v2_voc--------------------------------------------------------------------<br> https://mvsep.com/quality_checker/entry/9470<br> <code> Metric sdr for vocals: 11.3957<br> Metric si_sdr for vocals: 10.9629<br> Metric l1_freq for vocals: 39.5419<br> Metric log_wmse for vocals: 14.3935<br> Metric aura_stft for vocals: 10.2444<br> Metric aura_mrstft for vocals: 12.8111<br> Metric bleedless for vocals: 34.0758<br> Metric fullness for vocals: 19.0952<br> </code> <br> <br> v1--------------------------------------------------------------------<br> https://mvsep.com/quality_checker/entry/9303<br> <code> Metric sdr for instrum: 17.2702<br> Metric si_sdr for instrum: 17.1860<br> Metric l1_freq for instrum: 39.6100<br> Metric log_wmse for instrum: 13.9687<br> Metric aura_stft for instrum: 16.7484<br> Metric aura_mrstft for instrum: 22.7086<br> Metric bleedless for instrum: 38.7761<br> Metric fullness for instrum: 36.9074<br> </code>

To use this model, replace the model file with one from this repository.<br><br> Some components in the SegmModel module were implemented based on this paper: https://arxiv.org/abs/2506.17733.<br><br>

This weight is based on the following weights. Thank you, anvuew!<br> https://huggingface.co/anvuew/BS-RoFormer