audiosronnx-callenhancer

Pure-ONNX weights for the callenhancer engine in audiosronnx — call-centre / telephony speech restoration, 16 kHz → 48 kHz.

This is an ONNX export of Scicom-intl/CallEnhancer:

feature_extractor.int8.onnx — the full 24-layer w2v-BERT 2.0 feature predictor (LoRA merged into the base weights), int8-quantized for CPU. Maps SeamlessM4T log-mel features [1, T, 160] → hidden states [1, T, 1024] at 50 Hz.
decoder.onnx — the 188 M-param DAC vocoder (fp32), [1, 1024, T] → [1, 1, 960·T] 48 kHz waveform.

The SeamlessM4T log-mel front-end runs in numpy inside audiosronnx, so inference is onnxruntime-only (no torch). Each graph was validated against the original PyTorch model (feature extractor max abs err 1.9e-5, decoder 5.0e-6; numpy mel front-end 1.0e-5).

Usage

from audiosronnx import load_sr

sr = load_sr("callenhancer")
sr.upscale_file("call_8k.wav", "restored_48k.wav")

License

The model weights are CC-BY-NC-4.0 (research / non-commercial), inherited from the upstream CallEnhancer release, which builds on Meta's w2v-BERT 2.0 and the Descript Audio Codec. This covers the model itself — audio you restore with it is your own.

Downloads last month: -; Downloads are not tracked for this model. How to track