Struct AudioTokenizer 

pub struct AudioTokenizer {
    pub cfg: TokenizerConfig,
    /* private fields */
}

The MiMo audio tokenizer encoder and its residual VQ.

Fields

cfg: TokenizerConfig

Implementations

impl AudioTokenizer

pub fn features( &self, mel: &[f32], m: usize, pool: Option<&Pool>, ) -> Result<Vec<f32>, String>

Pre-RVQ features of ONE segment of m mel frames (mel [m][n_mels]): rows [⌈⌈m/2⌉/2⌉][d] after down_sample_norm.

pub fn quantize( &self, feats: &[f32], rows: usize, bf16_books: bool, pool: Option<&Pool>, ) -> Vec<u32>

Residual VQ in f32: per level, the nearest codeword by ‖E‖² − 2 r·E (the row’s ‖r‖² does not move the argmin), then r −= E[idx]. bf16_books selects the bf16-rounded codebooks a bf16 serving load holds. Returns [rows][levels].