Fabio
c36510c21e
Some checks failed
Linux Build / Linux cpu (push) Has been cancelled
Linux Build / Linux cpu-clang (push) Has been cancelled
Linux Build / Linux vulkan (push) Has been cancelled
mac-build / macOS CPU (push) Has been cancelled
Nix Build / Nix (cpu) (push) Has been cancelled
Nix Build / Nix (vulkan) (push) Has been cancelled
windows-build / Windows CPU (push) Has been cancelled
40 lines
1.3 KiB
Markdown
40 lines
1.3 KiB
Markdown
# Sidon
|
|
|
|
Sidon v0.1 restores single-speaker speech and produces mono audio at 48 kHz.
|
|
This is the original Sidon model, not DialogueSidon.
|
|
|
|
## Usage
|
|
|
|
```bash
|
|
audiocpp_cli --task s2s --family sidon \
|
|
--model /path/to/Sidon-GGUF/sidon-v0.1-f32.gguf \
|
|
--audio input.wav --backend cuda --threads 8 --log --out restored.wav
|
|
```
|
|
|
|
Stereo input is mixed to mono. The frontend normalizes the peak, applies a
|
|
50 Hz high-pass filter, and resamples to 16 kHz. Output is trimmed to the
|
|
original recording's duration.
|
|
|
|
Long recordings follow the released demo's 96-second chunks with a one-frame
|
|
encoder carry between chunks. This is offline processing, not streaming.
|
|
|
|
## Conversion
|
|
|
|
Download `feature_extractor_cuda.pt` and `decoder_cuda.pt` from
|
|
[sarulab-speech/sidon-v0.1](https://huggingface.co/sarulab-speech/sidon-v0.1).
|
|
Build `audiocpp_gguf`, then run:
|
|
|
|
```bash
|
|
python tests/sidon/convert_gguf.py \
|
|
--model-dir /path/to/sidon-v0.1 \
|
|
--output /path/to/Sidon-GGUF/sidon-v0.1-f32.gguf
|
|
```
|
|
|
|
The converter extracts the frozen TorchScript constants, folds the encoder's
|
|
LoRA adapters, and embeds the model spec. The resulting GGUF does not require
|
|
the original TorchScript files at runtime.
|
|
|
|
## Upstream
|
|
|
|
- [Model checkpoint](https://huggingface.co/sarulab-speech/sidon-v0.1)
|
|
- [Source and license](https://github.com/sarulab-speech/Sidon)
|