Sidon-GGUF

Sidon GGUF

GGUF quantization of SaruLab Sidon v0.1, a multilingual speech-restoration model that removes noise and reverberation and restores speech bandwidth. This repository packages the model for native inference with CrispASR.

The model contains the first eight layers of w2v-BERT 2.0 with the Sidon LoRA adapter merged, followed by Sidon's continuous DAC decoder. The exact SeamlessM4T log-mel window and filter bank are embedded in the GGUF.

Model

SHA-256:

7E27282A09BE71E2F37E87731A074B291A2CAB0D3F1DE4A8F8EB48421587E912  sidon-v0.1-f16.gguf

CrispASR usage

crispasr --s2s `
  -m sidon-v0.1-f16.gguf `
  -f input.wav `
  --s2s-output restored-48khz.wav

CrispASR detects the sidon architecture from GGUF metadata. The current S2S interface processes a complete clip and returns a complete restored clip; it is not a streaming interface.

GPU execution can be selected explicitly with --gpu-backend cuda or --gpu-backend vulkan. The Vulkan predictor uses native decompositions for operations that GGML's Vulkan backend does not currently fuse.

Conversion

python models/convert-sidon-to-gguf.py `
  --base w2v-bert-2.0 `
  --sidon sidon_raw_weight `
  --output sidon-v0.1-f16.gguf `
  --dtype f16

The converter consumes the original sidon_raw_weight adapter/decoder and facebook/w2v-bert-2.0 base weights.

Validation

On a one-second validation clip, native CrispASR output matched a PyTorch reconstruction of the same merged raw checkpoint at waveform cosine similarity 0.999966 for both the FP32 diagnostic conversion and this FP16 GGUF.

Sidon's separately published TorchScript bundle is a slightly different checkpoint revision, so direct comparison to that bundle is not an exact raw-checkpoint parity test.

The complete native pipeline was also tested on an NVIDIA GeForce RTX 5070 Ti with a 9.14-second input:

Vulkan-to-CPU waveform cosine was 0.999559. Backend timings include feature extraction, graph setup, inference, and output transfer; they are hardware- and driver-dependent.

License and attribution

Sidon, w2v-BERT 2.0, and the DAC decoder are released under the MIT license. See the Sidon repository and Sidon paper for authorship and training details.

Contributors

• Wataru Nakata
• Yuki Saito

Acknowledgements

The development of the original model was supported by project gamma of the National Institute of Advanced Industrial Science and Technology.