audiosronnx-sidon

ONNX weights for the Sidon speech-restoration engine of audiosronnx. Sidon restores degraded 16 kHz speech and resynthesises it at 48 kHz.

Two graphs:

• feature_extractor.int8.onnx — 8-layer w2v-BERT 2.0 feature predictor (LoRA-adapted), SeamlessM4T log-mel features [1, T, 160] → hidden states [1, T, 1024] at 50 Hz. Dynamic int8-quantized for CPU.
• decoder.onnx — DAC vocoder (960× upsample), [1, 1024, T] → [1, 1, 960·T] 48 kHz waveform. fp32.

The SeamlessM4T mel front-end runs in numpy inside the adapter, so inference is onnxruntime-only (no torch).

Converted from sarulab-speech/sidon-v0.1 (TorchScript). Graph parity vs. PyTorch: max abs err ≤ 3e-6.

Usage

from audiosronnx import load_sr
sr = load_sr("sidon")
sr.upscale_file("noisy_16k.wav", "restored_48k.wav")

Upstream: sarulab-speech/Sidon (MIT).