Verification: Is this checkpoint MVSep Entry #10146 ("unwa leap")?
Hi @pcunwa and community,
Could you confirm if the BS-Roformer-Leap checkpoint here is the exact same model evaluated under MVSep Entry #10146 ("unwa leap" - added on 2026-06-13)?
For reference, the MVSep #10146 metrics are:
Metrics (Instrumental):
- SDR: 17.7794 | SI-SDR: 17.6978
- L1 Freq: 39.9577 | Log WMSE: 14.5030
- Aura STFT: 16.0481 | Aura MRSTFT: 19.9063
- Bleedless: 43.1717 | Fullness: 31.1612
Metrics (Vocals):
- SDR: 11.7222 | SI-SDR: 11.3480
- L1 Freq: 39.8610 | Log WMSE: 14.7066
- Aura STFT: 10.0946 | Aura MRSTFT: 11.3655
- Bleedless: 38.9019 | Fullness: 17.0702
Also, for anyone running this locally, what are the recommended inference parameters (chunk size, overlap) for the best results?
Thanks!
The result for the standard Leap checkpoint was #10146, and for Xe it was #10178
Also, these test were conducted by anvuew.
Thanks a lot for that.
Thanks @pcunwa for the info, and amazing work on these models!
Hi @anvuew , since you ran the MVSep tests, could you please share the config.yaml and the exact inference parameters you used for both Leap and Leap Xe? I'd love to know your settings for chunk_size, overlap, segment_size... Thanks so much!
Thanks @pcunwa for the info, and amazing work on these models!
Hi @anvuew , since you ran the MVSep tests, could you please share the config.yaml and the exact inference parameters you used for both Leap and Leap Xe? I'd love to know your settings for chunk_size, overlap, segment_size... Thanks so much!
I only changed overlap to 4, Everything else stayed at the config values.