mazesmazes/libritts-mimi

Dataset with Mimi Codes

This dataset adds Mimi codec codes to parler-tts/libritts_r_filtered.

Dataset Description

Each sample contains:

• audio: Audio resampled to 24kHz (Mimi's native rate)
• codes: 8-layer Mimi codec codes (list of 8 lists of integers)
• text: Text transcription (from text_normalized column)
• Additional columns preserved from source dataset

Stats

• Source: parler-tts/libritts_r_filtered
• Splits: train.clean.360
• Samples: 112,326
• Audio Sample Rate: 24kHz
• Codec: Mimi (kyutai/mimi) with 8 codebooks

Usage

from datasets import load_dataset

ds = load_dataset("mazesmazes/libritts-mimi", split="train")

# Access audio and codes together
sample = ds[0]
audio = sample["audio"]  # {'array': [...], 'sampling_rate': 24000}
codes = sample["codes"]  # 8 lists of codec indices
text = sample["text"]

# Decode codes back to audio
import torch
from transformers import MimiModel

mimi = MimiModel.from_pretrained("kyutai/mimi")
codes_tensor = torch.tensor(codes).unsqueeze(0)  # (1, 8, seq_len)
with torch.no_grad():
    decoded = mimi.decode(codes_tensor)
    waveform = decoded.audio_values  # (1, 1, samples) at 24kHz

Source Dataset

• parler-tts/libritts_r_filtered

License

Same as source dataset.