UniverSR general-audio super-resolution implementation report

This report is for contributors who review UniverSR runtime admission, implementation parity, and validation evidence.

Decision

The official general-audio UniverSR profile is admitted as a separate native audio enhancement runtime. It supports speech, music, and sound effects with effective 8, 12, 16, or 24 kHz input bandwidth and always writes 48 kHz mono audio. It is described as audio super-resolution, not a complete mastering chain.

Source and license admission

The following table records the admitted source and checkpoint identities:

Item	Frozen identity	License/admission result
Source	woongzip1/UniverSR@26dc21c44e11f9f19e823f02b0d4641dd5ea5af2	MIT
Official checkpoint	woongzip1/universr-audio@1c3294844285af851b6ffa56cbde4e43cd41fc2b	CC BY 4.0

The source-code and checkpoint licenses are intentionally recorded separately. The 229,072,395-byte pytorch_model.bin has SHA-256 eb99f98943cc32fa82226c2da14b32b5d890416070af4946acbce442b30dc20b. The 533-byte source config.yaml and 1,406-byte official model card are also pinned by exact byte count and SHA-256. Runtime readiness fails if any of the three artifacts changes.

Native graph

The typed graph contains 394 float32 tensors and 57,231,302 scalars. It follows the published general-audio profile:

• Deterministic integer-ratio, windowed-sinc interpolation to 48 kHz
• Symmetric-Hann, 1,024-point complex short-time Fourier transform (STFT) with a 512-sample hop
• Power-law magnitude compression with alpha 0.2, beta 1, and epsilon 1e-4
• Conditional ConvNeXt V2 U-Net dimensions [96, 192, 384, 768], depths [2, 2, 4, 2], 384 conditioning channels, and a 256-dimensional time path
• Rate-conditioned low-frequency observations and reconstruction of the remaining spectrum
• Deterministic noise seeding and Euler, midpoint, or RK4 ordinary differential equation (ODE) integration
• Inverse compression and native inverse STFT cropped to the source duration

PyTorch Conv2d and ConvTranspose2d tensors are transformed into MLX kernel layouts only after the restricted state-dict reader verifies archive metadata, key set, shape, dtype, tensor count, scalar count, byte count, and whole-file hash.

Runtime boundary

mere.run audio enhance --model audio-enhance-universr-audio selects this runtime. Native 8/12/16/24 kHz files determine their effective rate directly; bandwidth-limited audio stored in a 48 kHz container requires --input-rate so the runtime does not guess. The published defaults are midpoint integration, four steps, guidance scale 1.5, and seed 42, all recorded in the output provenance manifest.

The installed-model gate and manual smoke use the public command to produce a WAV file. Finite, non-silent output and spectral energy above the declared input boundary prove the native path executed; they are not a perceptual quality benchmark.