SIREN-FIX

Neural Audio Restoration with Flow Matching

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

Model Description

SIREN-FIX restores and repairs damaged audio using a Flow Matching architecture with text-conditioned restoration. The model can denoise, declip, dereverb, and enhance audio quality guided by natural language prompts.

Key capabilities:

• Denoising - Remove background noise and hiss
• Declipping - Repair clipped/distorted audio
• Dereverberation - Remove unwanted room acoustics
• Text-guided restoration - Describe desired output in natural language
• General enhancement - Improve overall audio clarity

Architecture

Component	Details
Base Architecture	Flow Matching
Model Size	900M parameters
Sample Rate	44.1 kHz

The SIREN Family

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

Usage

import torch

# Load model
checkpoint = torch.load('siren_fix.pt', map_location='cpu')
model_state = checkpoint['model_state_dict']

# Model expects audio at 44.1kHz
# Input: damaged/noisy audio
# Output: restored clean audio

Training Details

• Training Data: Paired damaged/clean audio dataset
• Training Duration: 100 epochs
• Hardware: NVIDIA B200 GPUs

Intended Use

• Audio restoration and cleanup
• Noise removal from recordings
• Declipping distorted audio
• Removing unwanted reverb
• Research in neural audio restoration

Limitations

• Optimized for 44.1kHz sample rate
• Restoration quality depends on damage severity
• Text conditioning requires descriptive prompts

License

Apache 2.0

Citation

@software{siren_fix_2026,
  title={SIREN-FIX: Neural Audio Restoration with Flow Matching},
  author={SIREN Team},
  year={2026},
  url={https://huggingface.co/hilarl/siren-fix}
}