You need to agree to share your contact information to access this model

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

SIREN-SEPARATE

Two-Stage Source Separation with Query-Based Extraction

SIREN-SEPARATE is part of the SIREN Audio Suite - a family of neural audio processing models designed for professional music production workflows.

Model Description

SIREN-SEPARATE provides two-stage source separation. Stage 1 performs standard 4-stem separation. Stage 2 uses a fine-tuned query-conditioned model for granular extraction from the "Other" stem.

Key capabilities:

4-stem separation - Vocals, Drums, Bass, Other (Stage 1)
Query-based extraction - Extract specific instruments from Other (Stage 2)
Granular separation - Guitar, piano, synth, strings, any instrument
High quality - Fine-tuned from pretrained checkpoint

Architecture

Component	Details
Stage 1	Pretrained 4-stem separator (9.2 dB SDR)
Stage 2	Query-conditioned separator (fine-tuned)
Checkpoint Size	616MB
Sample Rate	44.1 kHz

The SIREN Family

Model	Purpose
SIREN-FX	Neural audio effects
SIREN-FIX	Audio restoration and repair
SIREN-MASTER	Audio enhancement and mastering
SIREN-STEER	Steerable audio transformations
SIREN-SEPARATE	Source separation (this model)
SIREN-TRANSCRIBE	Music analysis (key, tempo, beats)

Usage

import torch

# Load Stage 2 model (query-based separation)
checkpoint = torch.load('siren_separate.pt', map_location='cpu')

# Stage 1: Use pretrained separator for basic 4-stem separation
# Stage 2: Use this model for granular extraction from Other stem

Training Details

Training Data: Multi-track audio dataset
Hardware: NVIDIA B200 GPU
Fine-tuning: From pretrained separator checkpoint

Intended Use

Music source separation
Stem extraction for remixing
Instrument isolation
Karaoke (vocal removal)
Research in audio source separation

Limitations

Stage 1 limited to 4 stems (vocals, drums, bass, other)
Stage 2 quality depends on query audio similarity
Optimized for 44.1kHz sample rate

License

Apache 2.0

Citation

@software{siren_separate_2026,
  title={SIREN-SEPARATE: Two-Stage Source Separation},
  author={SIREN Team},
  year={2026},
  url={https://huggingface.co/hilarl/siren-separate}
}

Downloads last month: -; Downloads are not tracked for this model. How to track