AP-BWE speech bandwidth-extension implementation report

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

Decision

AP-BWE's 16 kHz to 48 kHz profile is admitted as a separate native audio enhancement runtime. It is intentionally not presented as music mastering or general audio super-resolution.

Source and license admission

The following table records the admitted source, checkpoint, and transport identities:

Item	Frozen identity	License/admission result
Source	yxlu-0102/AP-BWE@751710f22404c27e5bcc983248f8b856a04b8422	MIT
Official checkpoint	Google Drive file 1HYkD_5ha9GMrjzbTiSFiO1uQQwxXET15	Upstream states weights use the same MIT license
Managed transport	rsxdalv/AP-BWE@fa3f46d233cbc1d75cec9321188f86db627ba239	Public mirror used only for transport

The managed g_16kto48k.zip has 119,097,961 bytes and SHA-256 305a05dcab7dc29ffba09d32692d7a34550fc8fbdf338013641ff5d39a3cb285. Those bytes match the official Drive checkpoint exactly. The source config, code license, and weights-license files are also independently pinned by byte count and SHA-256. Runtime readiness fails if any of the four changes.

Native graph

The typed graph contains 162 float32 tensors and 29,760,515 scalars. It follows the published 16 kHz to 48 kHz profile:

• Mono 16 kHz decoding and deterministic threefold band-limited interpolation
• Centered 1,024-point short-time Fourier transform (STFT) with a periodic 320-sample Hann window, padding to the fast Fourier transform (FFT) size, and an 80-sample hop
• Magnitude and phase input convolutions from 513 bins to 512 channels
• Eight paired ConvNeXt blocks per stream with depthwise kernel-7 convolution, layer normalization, exact GELU, threefold channel expansion, layer scale, and residuals
• The upstream sequential magnitude and phase cross-add update
• Residual log-magnitude prediction, atan2 phase reconstruction, and native inverse STFT cropped to the input duration

PyTorch Conv1d tensors are transposed from [out, in, kernel] to the MLX [out, kernel, in] layout. The loader rejects key, shape, dtype, tensor-count, scalar-count, byte-count, and whole-file hash drift.

Runtime boundary

mere.run audio enhance is a top-level audio command so speech bandwidth extension and general-audio super-resolution do not get conflated with music stem separation. AP-BWE remains independently reviewable from the stacked UniverSR runtime.

The installed-model gate produces a WAV file through the public command. The manual checkpoint smoke also verifies finite, non-silent, 48 kHz mono output and reports energy above the original 8 kHz narrowband boundary. These checks prove execution and artifact integrity, not a perceptual benchmark.