Struct Mimi 

pub struct Mimi<T: Tensor> {
    pub quantizer: SplitResidualVectorQuantizer<T>,
    /* private fields */
}

Mimi audio tokenizer.

Fields

quantizer: SplitResidualVectorQuantizer<T>

Split residual vector quantizer.

Implementations

impl<T: Tensor> Mimi<T>

pub fn new(config: Config, context: &T::Context) -> Result<Self, Error>

Creates an unloaded Mimi tokenizer from config.

pub fn mimi_config(&self) -> &Config

Returns the Mimi configuration.

pub fn load_parameters( &mut self, parameters: impl IntoIterator<Item = (String, T)>, ) -> Result<(), Error>

Strictly replaces every model parameter with backend-native checkpoint tensors using Eredu’s stable parameter names.

Backend integrations remain responsible for reading their preferred checkpoint format and converting weight layouts. Mimi owns parameter completeness and shape validation, so those integrations do not need to reproduce the architecture or its parameter tree.

pub fn encode_latent( &mut self, latent: &T, context: &T::Context, ) -> Result<T, Error>

Encodes latent frames shaped [batch, 512, frames] into Mimi tokens.

pub fn encode(&mut self, pcm: &T, context: &T::Context) -> Result<T, Error>

Encodes PCM shaped [batch, 1, samples] into Mimi tokens [batch, codebooks, frames].

pub fn reset_encode_state(&mut self)

Resets state used by Mimi::encode_step.

pub fn encode_step( &mut self, pcm: &T, context: &T::Context, ) -> Result<Option<T>, Error>

Encodes one PCM frame into the next Mimi token frame.

Accepts PCM shaped [batch, 1, samples]. Returns None until the streaming encoder has enough samples to emit a complete codec frame.

pub fn decode_latent( &mut self, codes: &T, context: &T::Context, ) -> Result<T, Error>

Decodes Mimi tokens shaped [batch, codebooks, frames] into latent frames.

pub fn decode(&mut self, codes: &T, context: &T::Context) -> Result<T, Error>

Decodes Mimi tokens shaped [batch, codebooks, frames] into PCM [batch, 1, samples].

pub fn reset_decode_state(&mut self)

Resets state used by Mimi::decode_step.

pub fn decode_step( &mut self, codes: &T, context: &T::Context, ) -> Result<T, Error>

Decodes one Mimi token frame into the next PCM chunk.

Accepts codes shaped [batch, codebooks] or [batch, codebooks, 1].

Trait Implementations

impl<T: Tensor> AudioTokenizer for Mimi<T>

type Tensor = T

Backend-native tensor handle.

fn config(&self) -> AudioTokenizerConfig

Codec configuration.

fn encode(&mut self, pcm: &T, context: &T::Context) -> Result<T, Error>

Encodes mono PCM shaped [batch, channels, samples] into codec tokens.

fn decode(&mut self, codes: &T, context: &T::Context) -> Result<T, Error>

Decodes codec tokens shaped [batch, codebooks, frames] into PCM.

impl<T: Clone + Tensor> Clone for Mimi<T>

fn clone(&self) -> Mimi<T>

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl<T: Debug + Tensor> Debug for Mimi<T>

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl<T: Tensor> Parameterized<T> for Mimi<T>

fn visit_parameters<'a, V>(&'a self, visitor: &mut V)

where V: ParameterVisitor<'a, T>,

Visits every parameter exactly once using stable identities.

fn visit_parameters_mut<'a, V>(&'a mut self, visitor: &mut V)

where V: ParameterVisitorMut<'a, T>,

Mutably visits every parameter exactly once using stable identities.

fn set_trainable(&mut self, trainable: bool)

Updates whether all parameters in this module are trainable.