Apppublic

modulsx/bs-roformer-leap-audio-separator

sourceHugging Faceupdated 2mo agoView on Hugging Face

0likes

App README

BS-Roformer Leap Audio Separator

This Space demonstrates audio source separation using the BS-Roformer Leap model checkpoints.

Upload any audio file and choose one of four model variants to extract vocals or instruments:

Variant	Target	Config	Checkpoint
Standard — Vocals	vocals	`bs_leap_voc_conf.yaml`	`bs_roformer_leap_voc.ckpt`
Standard — Instruments	instruments	`bs_leap_inst_conf.yaml`	`bs_roformer_leap_inst.ckpt`
Xe — Vocals	vocals	`Xe/leap_xe_config_voc.yaml`	`Xe/bs_leap_xe_voc.ckpt`
Xe — Instruments	instruments	`Xe/leap_xe_config_inst.yaml`	`Xe/bs_leap_xe_inst.ckpt`

Architecture

BS-Roformer (Band-Split Roformer) is a transformer-based architecture for audio source separation. It divides the frequency spectrum into bands and processes each band with a Roformer (rotary positional encoding transformer).

Key config: dim=256, depth=16, 8 heads, dim_head=64, stereo, 44.1 kHz, flash attention (PyTorch SDPA).

Inference

The model processes audio in overlapping chunks using overlap-add with a Hamming window, matching the original UVR5 inference pipeline.