jspaulsen/unmute-encoder

Unmute Encoder

A speaker embedding encoder trained to replicate Kyutai's unreleased "unmute encoder". This model extracts speaker embeddings from audio for use with Kyutai's Moshi TTS system.

Model Description

The encoder is built on top of Kyutai's Mimi neural audio codec:

1. 1.Mimi Encoder: Frozen Mimi encoder extracts latent audio representations
2. 2.MLP Projector: Trainable MLP head projects Mimi's latents to the target embedding space
3. 3.Output: Speaker embeddings of shape [512, 125] (512 channels, 125 time steps for 10s audio)

Audio (24kHz, 10s) -> Mimi Encoder -> Latent [512, T] -> MLP Projector -> Embedding [512, 125]

Usage

from src.models.mimi import MimiEncoder

# Load the encoder
encoder = MimiEncoder.from_pretrained(
    model_name="jspaulsen/unmute-encoder",
    device="cuda",
    num_codebooks=32,
)

# Create embedding from audio tensor [1, 1, T] at 24kHz
output = encoder(audio_tensor)
embedding = output.embedding  # [1, 512, 125]

Training

Trained using supervised learning with a hybrid loss (L1 + cosine similarity) against speaker embeddings from kyutai/tts-voices.

Training Details

• —Global step: 500
• —Epoch: 83.33333333333333
• —Best metric: 0.17919586598873138

Acknowledgments

• —Kyutai for releasing the Moshi TTS models and speaker embeddings