Datasets:

shangeth
/

librispeech-mimi-codes

license: cc-by-4.0
language:
  - en
task_categories:
  - text-to-speech
  - automatic-speech-recognition
tags:
  - mimi
  - neural-codec
  - speech-synthesis
  - speech-recognition
  - librispeech
  - audio-tokens
pretty_name: LibriSpeech Mimi Codes
size_categories:
  - 100K<n<1M
configs:
  - config_name: default
    data_files:
      - split: dev_clean
        path: data/dev_clean-*
      - split: dev_other
        path: data/dev_other-*
      - split: test_clean
        path: data/test_clean-*
      - split: test_other
        path: data/test_other-*
dataset_info:
  features:
    - name: id
      dtype: string
    - name: text
      dtype: string
    - name: speaker_id
      dtype: int32
    - name: codes
      list:
        list: int16
    - name: n_frames
      dtype: int32
    - name: k_codebooks
      dtype: int32
  splits:
    - name: dev_clean
      num_bytes: 4381124
      num_examples: 2703
    - name: dev_other
      num_bytes: 4178986
      num_examples: 2864
    - name: test_clean
      num_bytes: 4377566
      num_examples: 2620
    - name: test_other
      num_bytes: 4350511
      num_examples: 2939
  download_size: 11547138
  dataset_size: 17288187

LibriSpeech — Mimi Codes

Pre-extracted Kyutai Mimi neural-codec tokens for the LibriSpeech corpus — multi-speaker English audiobook readings from the LibriVox project.

This dataset contains codes only, not audio. For waveforms, use any of the LibriSpeech mirrors (e.g. openslr/librispeech_asr); these codes let you skip the ~hours of GPU extraction needed to train Mimi-based speech models.

Schema

One row per utterance:

Column	Type	Notes
`id`	string	`{speaker_id}-{chapter_id}-{utterance_id:04d}`, e.g. `103-1240-0000`
`text`	string	lowercased transcript
`speaker_id`	int32	LibriSpeech speaker ID
`codes`	`int16[k=8][n_frames]`	Mimi codebook indices @ 12.5 fps
`n_frames`	int32	= `codes.shape[1]`
`k_codebooks`	int32	= 8

Extraction details

Codec: kyutai/mimi @ 24 kHz, 12.5 fps
Codebooks: all 8 extracted. Slice codes[:k] for fewer (Mimi's codebooks are ordered by importance; the first few capture most of the signal).
Codebook size: 2048 per codebook → values stored as int16
Transcripts: sourced from LibriSpeech's .trans.txt files, lowercased (the raw release is ALL-UPPER)

Splits

Each standard LibriSpeech split is a separate HF split (hyphens replaced with underscores):

HF Split	Upstream	Approx. rows	Notes
`train_clean_100`	`train-clean-100`	~28.5k	clean read speech, ~100 h
`train_clean_360`	`train-clean-360`	~104.0k	clean read speech, ~360 h
`train_other_500`	`train-other-500`	~148.7k	noisier/accented, ~500 h
`dev_clean`	`dev-clean`	~2.7k	dev set, clean
`dev_other`	`dev-other`	~2.9k	dev set, noisier
`test_clean`	`test-clean`	~2.6k	test set, clean
`test_other`	`test-other`	~2.9k	test set, noisier

Only splits the publisher had on disk are included — consult the "Files" tab for the exact subset in this release.

Usage

from datasets import load_dataset
import torch

ds = load_dataset("shangeth/librispeech-mimi-codes", split="train_clean_100")

ex    = ds[0]
codes = torch.tensor(ex["codes"], dtype=torch.long)   # [8, n_frames]
print(f"{ex['id']} (speaker {ex['speaker_id']}) → {ex['text'][:60]}")
print("codes:", codes.shape, "duration:", codes.shape[1] / 12.5, "s")

# Use only the first 3 codebooks:
codes_3 = codes[:3]

Streaming (no full download):

ds = load_dataset("shangeth/librispeech-mimi-codes", split="train_clean_360", streaming=True)
for ex in ds.take(10):
    print(ex["id"], len(ex["codes"]), "codebooks")

Decode to audio with the Mimi decoder:

from transformers import MimiModel
mimi = MimiModel.from_pretrained("kyutai/mimi").cuda().eval()
with torch.no_grad():
    wav = mimi.decode(codes.unsqueeze(0).cuda()).audio_values[0].cpu()
# wav is [1, T] @ 24 kHz

License & Attribution

LibriSpeech is released under CC-BY-4.0. The derived Mimi codes inherit this license — attribution is required. Please cite both the original corpus and this dataset when redistributing.

Citations

@inproceedings{panayotov2015librispeech,
  title     = {Librispeech: an ASR corpus based on public domain audio books},
  author    = {Panayotov, Vassil and Chen, Guoguo and Povey, Daniel and Khudanpur, Sanjeev},
  booktitle = {ICASSP},
  year      = {2015}
}

@article{defossez2024moshi,
  title   = {Moshi: a speech-text foundation model for real-time dialogue},
  author  = {D{\'e}fossez, Alexandre and others},
  year    = {2024}
}

Used to train the Wren series of small speech LLMs.