File size: 1,050 Bytes
258647b | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 | # SNAC 24 kHz decoder (ONNX)
The decoder of [SNAC 24 kHz](https://huggingface.co/hubertsiuzdak/snac_24khz) (revision `d73ad176a12188fcf4f360ba3bf2c2fbbe8f58ec`, MIT licence, by Hubert
Siuzdak), exported to ONNX for torch-free decoding of the audio codes that Orpheus-style text-to-speech models
(Orpheus, Veena, Svara, snorTTS) generate. Exported with `tools/export_snac_onnx.py` from the bolo speech server.
| file | precision | size |
| --- | --- | --- |
| `snac_24khz_decoder.onnx` | fp32 | 52.6 MB |
| `snac_24khz_decoder_fp16.onnx` | fp16 | 26.4 MB |
Inputs: `codes_0` `[B, T]`, `codes_1` `[B, 2T]`, `codes_2` `[B, 4T]` (int64, 0..4095). Output: `audio`
`[B, 1, 2048·T]`, float32 at 24 kHz. The noise blocks of the decoder are kept (random, as in PyTorch).
Checked against PyTorch SNAC with the noise term disabled on both sides: see `snac_onnx.json`. Use fp32 or fp16;
int8 quantisation of this decoder is known to fail (audible noise), so none is provided.
Licence: MIT, as the original SNAC weights (Copyright (c) 2024 Hubert Siuzdak).
|