Module mimi 

Mimi neural audio tokenizer support. Mimi neural audio tokenizer support.

Mimi is the neural audio codec used by Moshi-family realtime models. This module implements backend-neutral checkpoint parameters, the split residual vector quantizer, and the non-streaming SEANet/transformer encoder and decoder used to map between PCM and Mimi codebook tokens.

Structs

CheckpointTensorPlan

Backend-independent loading plan for one tensor in a released Mimi checkpoint.

Config

Mimi codec configuration.

Conv1x1NoBias

Bias-free 1x1 convolution over [batch, channels, frames] tensors.

EuclideanCodebook

Euclidean codebook backed by EMA cluster statistics.

Mimi

Mimi audio tokenizer.

ResidualVectorQuantization

Residual vector quantization layers.

ResidualVectorQuantizer

Residual vector quantizer branch.

SplitResidualVectorQuantizer

Split residual vector quantizer used by Mimi.

VectorQuantization

Single vector-quantization layer.

Enums

CheckpointTensorLayout

Backend-independent layout conversion required by a Mimi checkpoint tensor.

ResampleMethod

Mimi resampling strategy.

Functions

checkpoint_tensor_plan

Maps a released Mimi checkpoint tensor name to its stable model parameter and required layout conversion.