eredu-codec

eredu-codec contains backend-neutral neural-audio architectures used with Eredu's realtime speech models. Backends construct Mimi<T>, use checkpoint_tensor_plan to map released checkpoint tensors, and populate the model through Mimi::load_parameters.

Mimi

The mimi module implements the Mimi encoder, residual vector quantizer, and decoder used by Moshi-family speech models. It supports:

• backend-neutral checkpoint name and tensor-layout planning;
• selecting an active subset of a checkpoint's codebooks;
• PCM-to-token and token-to-PCM conversion;
• latent-to-token and token-to-latent conversion; and
• stateful one-frame decoding for realtime playback.

use eredu_codec::mimi::Mimi;
use eredu_nn::Tensor;

fn round_trip<T: Tensor>(
    mimi: &mut Mimi<T>,
    pcm: &T,
    context: &T::Context,
) -> Result<T, eredu_codec::Error> {
    let tokens = mimi.encode(pcm, context)?;
    mimi.decode(&tokens, context)
}

Tensor shapes follow [batch, channels, samples_or_frames]. Audio capture, playback, resampling, and device selection remain application concerns.

Evaluation tools

Concrete backend integrations own executable Mimi benchmarks and PersonaPlex evaluation entry points; this crate has no concrete-backend feature or accelerator dependency. See the PersonaPlex quantization evaluation guide.

License

Licensed under either Apache-2.0 or MIT.