SNAC 24 kHz decoder (ONNX)

The decoder of SNAC 24 kHz (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.