AP-BWE β€” ONNX (8 kHz and 12 kHz to 48 kHz, fixed 21 s window)

ONNX exports of the generator of AP-BWE (Lu, Ai, Du, Ling: Towards High-Quality and Efficient Speech Bandwidth Extension with Parallel Amplitude and Phase Prediction, code), from the weights mirrored at rsxdalv/AP-BWE (revision fa3f46d2…).

Exported for the native (tract, pure Rust) enhancer runtime of cleanin, a speech cleanup tool. Registry id there: ap-bwe-native.

file checkpoint sha256
ap_bwe_8k_48k.onnx 8kto48k dc147c930f135e1f78bd74fa76ad182882e3c7fc5cca2e5d273e00ebe61a344d
ap_bwe_12k_48k.onnx 12kto48k 12a0613728582d04f05a49d91a64d2811ee2a59df1024b92777cc2e3bea6f02c

Each is 119 105 097 bytes, opset 17, f32 weights, static shapes for a 21 s window at 48 kHz (n_fft 1024, hop 80): inputs mag_nb = log(|Z| + 1e-4) and pha_nb = angle(Z) of the sinc-upsampled narrowband signal, [1, 513, 12601]; outputs mag_wb (log magnitude) and pha_wb, same shape. The STFT, band merge and iSTFT are not in the graph (cleanin does them in Rust). Max relative error against the original PyTorch generator: 3.7e-5. Details in MANIFEST.json. Export script: scripts/export-ap-bwe.py in the cleanin repository.

The 16 kHz and 24 kHz checkpoints are not exported here.

License

MIT, the same as the original code and weights. This repository contains a format conversion; all credit goes to the AP-BWE authors.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Model tree for mhavo/ap-bwe-onnx

Base model

rsxdalv/AP-BWE
Quantized
(1)
this model