universr-audio

UniverSR - General Audio (Native SafeTensors)

Vocoder-free broadband audio super-resolution model that upsamples 8 / 12 / 16 / 24 kHz → 48 kHz audio using Continuous Normalizing Flows (Flow Matching ODEs) in the power-compressed complex STFT domain (|X|⁰·² · e^(j∠X)). Trained on diverse acoustic distributions spanning speech, polyphonic music, and studio sound effects.

This repository provides verified, zero-overhead SafeTensors weights (model.safetensors) converted from the original checkpoint published by Choi et al., eliminating unsafe PyTorch pickle deserialization (pytorch_model.bin) while enabling instant zero-copy memory mapping. Integrated and optimized for production audio restoration within Furgie V2.

• Paper: arXiv:2510.00771
• Original Architecture: github.com/woongzip1/UniverSR
• Furgie Production Engine: github.com/oldskool978/Furgie

---

Validated Optimal Inference Defaults

• Target Delivery Mode: 48k (48.0 kHz 32-bit Float PCM Master)
• Flow ODE Solver: res_multistep_cfg_pp (2nd-Order Adams-Bashforth with Terminal Trapezoidal Anchoring)
• Integration Steps (N): 16 steps (17 total NFEs; 1.87x faster than Heun baseline)
• Guidance Scale (w): 0.00 (Single-pass conditional vector field evaluation)
• Deterministic Prior: seed=42 (Stateless hardware-accelerated Philox-4x32 PRNG)
• Crossover Passband Splice: 0 bins (Bit-exact ground-truth passband concatenation, 0.0 Hz - 12.0 kHz)
• Conditioning Anchor: 24000 Hz (Dynamic spectral partitioning for 8k, 12k, 16k, and 24k anchors)
• Headroom Strategy: bypass (Passband bit-exact unity, 1.0x linear gain scalar)

---

Usage

Via Furgie Production Harness

git clone https://github.com/oldskool978/Furgie.git
cd Furgie
pip install -r requirements.txt

# Download and verify SafeTensors weights
python scripts/hydrate_models.py --precision fp32

# Launch interactive super-resolution harness
python harness.py

# Non-interactive CLI batch generation
python harness.py --batch --input "workspace/input/track.wav" --output "workspace/output/master_48k.wav" --solver res_multistep_cfg_pp --steps 16 --cfg 0.0 --anchor 24000 --seed 42 --headroom-mode bypass

Python API Integration

import torch
from furgie_core.engine import FurgieEngine
from furgie_core.schema import FurgieRequest

# Initialize neural engine
engine = FurgieEngine(
    device="cuda" if torch.cuda.is_available() else "cpu",
    model_repo_id="OLDSKOOL978/universr-audio"
)

# Configure restoration request
request = FurgieRequest(
    input_path="workspace/input/track.wav",
    output_path="workspace/output/master_Furgie_48k.wav",
    solver="res_multistep_cfg_pp",
    ode_steps=16,
    guidance_scale=0.0,
    seed=42,
    input_sr_anchor=24000,
    headroom_mode="bypass",
    target_rate="48k"
)

# Run super-resolution pass
telemetry = engine.synthesize_request(request)
print(f"Synthesis Complete: {telemetry.duration_seconds:.2f}s audio generated in {telemetry.generation_time_seconds:.2f}s (RTF: {telemetry.real_time_factor:.3f}x)")
print(f"Phase Curvature: {telemetry.crossover_phase_delta_rad:.4f} rad | Crossover Step: {telemetry.crossover_magnitude_step_db:.3f} dB")

---

Architectural Configuration (config.yaml)

model_metadata:
  name: Furgie-Convergent-48K
  version: 2.0.0
  description: Filterless Complex STFT Generative Super-Resolution Engine.

audio:
  target_sample_rate: 48000
  n_fft: 1024
  hop_length: 512
  win_length: 1024
  power_alpha: 0.2

model:
  dims: [96, 192, 384, 768]
  depths: [2, 2, 4, 2]
  time_dim: 256
  cond_dim: 384
  total_freq_bins: 512
  hr_freq_bins: 432
  feature_enc_layers: 4
  sr_to_lr_bins:
    8: 80
    12: 128
    16: 170
    24: 256

universr_flow_core:
  enabled: true
  repo_id: OLDSKOOL978/universr-audio
  solver: res_multistep_cfg_pp
  ode_steps: 16
  guidance_scale: 0.0
  scheduler_type: uniform
  time_warp_gamma: 1.0
  seed: 42
  crossover_blend_bins: 0
  input_sr_anchor: 24000

---

Citations

Original UniverSR Architecture

@inproceedings{choi2026universr,
  title     = {{UniverSR}: Unified and Versatile Audio Super-Resolution via Vocoder-Free Flow Matching},
  author    = {Choi, Woongjib and Lee, Sangmin and Lim, Hyungseob and Kang, Hong-Goo},
  booktitle = {IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP)},
  year      = {2026}
}

Furgie SafeTensors & Super-Resolution Harness

@software{furgie2026,
  author    = {OLDSKOOL978},
  title     = {Furgie: Optimal Transport Flow-Matching Audio Super-Resolution Harness},
  url       = {https://github.com/oldskool978/Furgie},
  year      = {2026}
}