mira-iitjmu/svara-mimi-indic-v3

Svara Mimi Indic v3

Mimi audio codec fine-tuned on 10 Indian languages (IndicVoices-R).

Training

• Base model: kyutai/mimi
• Fine-tuned: encoder + semantic codebook (codebook 0)
• Decoder: frozen
• Loss: multi-scale mel reconstruction + WavLM distillation
• Data: IndicVoices-R, 10k samples/language, 10 languages
• Epochs: 10, batch_size: 8, lr: 1e-4

Usage

from svara.codec.mimi import MimiCodec
import torch

codec = MimiCodec.from_pretrained("kyutai/mimi", dtype=torch.bfloat16)
state = torch.load("mimi_final.pt", map_location="cpu")
codec.model.load_state_dict(state, strict=False)

Files

• mimi_final.pt — final checkpoint (full model state_dict)
• mimi_step*.pt — intermediate checkpoints