Audio enhancement runtime

Extend narrowband speech or bandwidth-limited general audio to a 48 kHz artifact with native Swift/MLX ports of AP-BWE and UniverSR. Both runtimes are local-only, verify their pinned artifacts before loading, and keep machine-readable results on stdout.

Command

Command	What it does
mere.run audio enhance	Reconstruct a 48 kHz mono float WAV with AP-BWE speech bandwidth extension or UniverSR general-audio super-resolution.

Models

• audio-enhance-ap-bwe-16kto48k
• audio-enhance-universr-audio

The source implementation is pinned to yxlu-0102/AP-BWE@751710f. The official repository states that both code and pretrained weights are MIT. The public Hugging Face transport snapshot is pinned independently. mere.run accepts its 16 to 48 kHz archive only because it is byte-identical to the official Google Drive checkpoint.

UniverSR's native graph is pinned to woongzip1/UniverSR@26dc21c, whose code is MIT-licensed. Its official woongzip1/universr-audio@1c32948 checkpoint is separately licensed CC BY 4.0. The managed profile preserves that distinction and pins the checkpoint, source configuration, and model card independently.

Typical workflow

swift run mere.run model pull audio-enhance-ap-bwe-16kto48k
swift run mere.run audio enhance ./narrowband-speech.wav \
  --output ./wideband-speech.wav

swift run mere.run model pull audio-enhance-universr-audio
swift run mere.run audio enhance ./music-12k.wav \
  --model audio-enhance-universr-audio \
  --output ./music-super-resolved.wav

# Declare the effective bandwidth when a limited source is stored at 48 kHz.
swift run mere.run audio enhance ./limited-48k.wav \
  --model audio-enhance-universr-audio \
  --input-rate 16000 --seed 42

Input is decoded to 16 kHz mono, then upsampled threefold with a deterministic windowed-sinc/Lanczos filter before the neural reconstruction stage. The pinned profile uses a 1,024-point short-time Fourier transform (STFT), 320-sample periodic Hann window, 80-sample hop, 512 channels, and eight paired magnitude/phase ConvNeXt blocks. Chunked overlap uses two-second 48 kHz windows by default.

UniverSR accepts effective input rates of 8, 12, 16, or 24 kHz. It applies a 1,024-point complex STFT, conditions a 394-tensor ConvNeXt V2 U-Net on the observed low-frequency region, and reconstructs the remaining spectrum with flow matching. The published defaults are midpoint integration, four ordinary differential equation (ODE) steps, classifier-free guidance 1.5, and seed 42. Native-rate files supply the effective rate automatically; a bandwidth-limited 48 kHz container requires --input-rate.

The command writes the float WAV and a sibling JSON file. That manifest records the source and output hashes, original and decoded audio geometry, immutable code and artifact revisions, checkpoint/config hashes, license identities, compute type, inference controls, chunk count, and elapsed time. The same JSON is written to stdout; progress remains on stderr.

AP-BWE is a speech bandwidth-extension model, not a general mastering model. UniverSR covers speech, music, and sound effects, but is still super-resolution rather than a complete mastering chain. A decoded, finite, non-silent output proves the native route executed; it does not by itself establish perceptual improvement for a particular recording.

See the AP-BWE implementation report for admission evidence and the UniverSR implementation report for its separate graph, artifact, and license evidence.