Datasets:

shangeth
/

mls-mimi-codes

license: cc-by-4.0
language:
  - nl
  - fr
  - de
  - it
  - pl
  - pt
  - es
task_categories:
  - text-to-speech
  - automatic-speech-recognition
tags:
  - mimi
  - neural-codec
  - multilingual
  - mls
  - multilingual-librispeech
  - audio-tokens
pretty_name: Multilingual LibriSpeech Mimi Codes
size_categories:
  - 1M<n<10M

Multilingual LibriSpeech (MLS) — Mimi Codes

Pre-extracted Kyutai Mimi neural-codec tokens for Multilingual LibriSpeech — LibriVox audiobooks in 7 non-English languages.

English is intentionally excluded. For English Mimi codes, use:

shangeth/librispeech-mimi-codes — LibriSpeech (~280k rows, 7 splits)
shangeth/libritts-r-mimi-codes — LibriTTS-R (~360k rows, 7 splits, 24 kHz native)
shangeth/vctk-mimi-codes — VCTK (~44k rows, 110 speakers w/ accents)
shangeth/jenny-mimi-codes — Jenny TTS (~21k rows, single speaker)
shangeth/ljspeech-mimi-codes — LJSpeech (~13k rows, single speaker)

Configs (languages)

One HF dataset config per language:

Config	Language	ISO	Approx hours (train)
`dutch`	Dutch	nl	~1.5k
`french`	French	fr	~1.1k
`german`	German	de	~3.3k
`italian`	Italian	it	~250
`polish`	Polish	pl	~100
`portuguese`	Portuguese	pt	~160
`spanish`	Spanish	es	~920

Splits (per config)

Split	Description
`train`	full training set
`dev`	development
`test`	test
`9_hours`	low-resource ~9h training subset
`1_hours`	low-resource ~1h training subset

Merged `all` config

A virtual all config aliases the per-language parquets via glob — no extra storage, just a YAML entry. Use it to train a single multilingual model:

ds = load_dataset("shangeth/mls-mimi-codes", "all", split="train")

Caveat: there is no language column, and speaker_id is per-language — so speaker IDs may collide across languages (e.g. German speaker 12345 is unrelated to French speaker 12345). If you need clean per-language speaker bookkeeping, load each language config separately.

Schema

Column	Type	Notes
`id`	string	utterance ID, format `{speaker}_{chapter}_{segment}`
`text`	string	transcript, mixed-case as-is from MLS
`speaker_id`	int32	speaker ID (parsed from MLS string)
`chapter_id`	int32	chapter ID
`codes`	`int16[k=8][n_frames]`	Mimi codebook indices @ 12.5 fps
`n_frames`	int32
`k_codebooks`	int32	8

Extraction details

Codec: kyutai/mimi @ 24 kHz, 12.5 fps
Resampling: MLS audio is 48 kHz opus → resampled to 24 kHz at extraction
Codebooks: all 8 extracted; slice codes[:k] for fewer
Source: facebook/multilingual_librispeech

Usage

from datasets import load_dataset
import torch

ds = load_dataset("shangeth/mls-mimi-codes", "german", split="dev")
ex = ds[0]
codes = torch.tensor(ex["codes"], dtype=torch.long)  # [8, n_frames]
print(ex["id"], "| speaker:", ex["speaker_id"], "|", ex["text"][:60])

# Decode back to 24 kHz audio
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()

Citation

@misc{wren2026,
  title  = {Wren: A Family of Small Open-Weight Models for Unified Speech-Text Modelling},
  author = {Shangeth Rajaa},
  year   = {2026},
  url    = {https://github.com/shangeth/wren}
}

@inproceedings{pratap2020mls,
  title     = {MLS: A Large-Scale Multilingual Dataset for Speech Research},
  author    = {Pratap, Vineel and Xu, Qiantong and Sriram, Anuroop and Synnaeve, Gabriel and Collobert, Ronan},
  booktitle = {Interspeech},
  year      = {2020}
}

License

CC-BY-4.0 (inherited from MLS / LibriVox).