Struct NeuCodecDecoder 

pub struct NeuCodecDecoder { /* private fields */ }

NeuCodec decoder: converts speech token IDs to a 24 kHz audio waveform.

Setup

Set NEUTTS_DECODER_PATH to neucodec_decoder.safetensors, then:

let dec = NeuCodecDecoder::new()?;
let audio = dec.decode(&codes)?;

Backend selection

When built with --features wgpu, the decoder automatically selects the best available backend on the first call to [decode] (lazy init):

Priority	Backend	When used
1	Burn wgpu (GPU)	Metal / Vulkan / DX12 adapter found
2	Burn NdArray (CPU)	No GPU adapter available
3	Raw ndarray (CPU)	Burn init failed entirely

The Burn backend is initialised eagerly at from_file time so the GPU upload cost is part of model loading, not synthesis latency.

Implementations

impl NeuCodecDecoder

pub fn new() -> Result<NeuCodecDecoder, Error>

Load from NEUTTS_DECODER_PATH.

pub fn from_file(path: &Path) -> Result<NeuCodecDecoder, Error>

Load from an explicit file path.

pub fn decode(&self, codes: &[i32]) -> Result<Vec<f32>, Error>

Decode speech token IDs to a 24 kHz audio waveform.

• codes — integer token IDs in 0..=65535 (NeuCodec FSQ range). Out-of-range values are rejected with an error rather than silently producing garbage digits from the FSQ decomposition.
• returns — Vec<f32> of codes.len() × hop_length samples.

pub fn backend_name(&self) -> &'static str

Name of the active inference backend.

pub fn load(path: &Path) -> Result<NeuCodecDecoder, Error>

Alias for from_file — load from an explicit path.

pub fn weights_path(&self) -> &Path

Path from which the decoder was loaded.

pub fn hop_length(&self) -> usize

Detected hop_length (audio samples per speech token).