Hugging Face's logo

Datasets:
shangeth
/
expresso-mimi-codes

Tasks:
Text-to-Speech
Automatic Speech Recognition
Modalities:
Tabular
Text
Formats:
parquet
Languages:
English
Size:
10K - 100K
ArXiv:
Tags:
mimi
neural-codec
speech-synthesis
expresso
audio-tokens
expressive-speech
Libraries:
Datasets
pandas
Polars
License:
Dataset card Data Studio Files
xet
 Community
expresso-mimi-codes / README.md
shangeth's picture
shangeth
Update dataset card
a25dcf3 verified
preview code
|
raw
history blame contribute delete
6.27 kB
metadata 
license: cc-by-nc-4.0
language:
  - en
task_categories:
  - text-to-speech
  - automatic-speech-recognition
tags:
  - mimi
  - neural-codec
  - speech-synthesis
  - expresso
  - audio-tokens
  - expressive-speech
  - style-transfer
  - disentanglement
pretty_name: Expresso — Mimi Codes (k=32)
size_categories:
  - 10K<n<100K
configs:
  - config_name: read
    data_files:
      - split: train
        path: read/train-*
      - split: dev
        path: read/dev-*
      - split: test
        path: read/test-*
  - config_name: conversational
    data_files:
      - split: train
        path: conversational/train-*
      - split: dev
        path: conversational/dev-*
      - split: test
        path: conversational/test-*

Expresso — Mimi Codes (k = 32)

Pre-extracted Kyutai Mimi tokens (all 32 codebooks) for both the read and conversational subsets of Expresso. Source audio + transcripts live in shangeth/expresso; this dataset publishes the discrete-token version for training Mimi-based speech models without re-extracting.

⚠️ License: CC-BY-NC-4.0 — non-commercial use only.

Why Expresso for Wren?

Expresso is the most directly relevant dataset for speech disentanglement research — the same speakers utter similar content across different expressive styles (happy, sad, confused, whispered, animal, etc.). Useful for studying:

  • Style vs content vs speaker identity
  • How expressive variation is encoded across Mimi's 32 codebooks (semantic + acoustic hierarchy)
  • Controllable style transfer at fixed speaker identity

Configs

  • read — ~11.6k mono utterances with human transcripts.
  • conversational — ~15.9k mono per-utterance turns from stereo dialogues, transcribed with Whisper Large V3 Turbo (3.0% overall WER on a held-out human-transcribed set).

Why 32 codebooks?

Most published speech-codec datasets use only the first 8 of Mimi's 32 codebooks (1 semantic + 7 acoustic), which is enough for the original Moshi recipe. We publish all 32 so you can:

  • Train models on more codebooks for higher resynthesis fidelity
  • Study which codebooks carry which content (style/timbre/prosody)
  • Slice codes[:k] at load time to use any prefix

k_codebooks is stored per row so the schema works for both k=32 and any subset you slice.

Schemas

read

Column Type Notes
id string e.g. ex01_confused_00001; longform: ex01_default_longform_00001 (full file in train)
text string human transcription (mixed case + punctuation); empty for chunked rows
speaker_id int32 1–4
style string default, confused, enunciated, happy, laughing, narration, sad, whisper
substyle string finer label (e.g. default_emphasis, default_essentials, default_longform)
corpus string base or longform
start_s / end_s float32 null for full-file rows
codes int16[32][n_frames] Mimi codebook indices @ 12.5 fps
n_frames int32
k_codebooks int32 32

conversational

Column Type Notes
id string e.g. ex01-ex02_default_001__ch1_23.88-28.14
text string Whisper Large V3 Turbo transcript
speaker_id int32 this channel's speaker (1–4)
style string this channel's style
other_speaker_id int32 partner's speaker id
other_style string partner's style
source_file_id string the parent stereo file
channel int32 1 or 2
start_s / end_s float32 turn boundaries within source file
codes int16[32][n_frames] Mimi codebook indices @ 12.5 fps
n_frames int32
k_codebooks int32 32

Extraction details

  • Source audio: shangeth/expresso (the official Expresso tar, segmented and built into HF format)
  • Codec: kyutai/mimi @ 24 kHz, 12.5 fps, codebook size 2048 (fits int16)
  • Resampling: 48 kHz mono → 24 kHz before encoding
  • Conversational text: machine-transcribed (Whisper Large V3 Turbo with anti-hallucination decoding)

Usage 

from datasets import load_dataset
import torch

# Pick a config
read = load_dataset("shangeth/expresso-mimi-codes", "read",          split="train")
conv = load_dataset("shangeth/expresso-mimi-codes", "conversational", split="train")

ex = read[0]
codes = torch.tensor(ex["codes"], dtype=torch.long)  # [32, n_frames]
print(f"speaker={ex['speaker_id']} style={ex['style']} | {ex['text'][:60]}")
print(f"codes shape: {codes.shape}  ({codes.shape[1]/12.5:.2f}s @ 12.5 fps)")

# Use only the first 8 codebooks (Moshi-style)
codes8 = codes[:8]

# 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()  # [1, T] @ 24 kHz

Links

Citation 

@inproceedings{nguyen2023expresso,
  title     = {Expresso: A Benchmark and Analysis of Discrete Expressive Speech Resynthesis},
  author    = {Nguyen, Tu Anh and Hsu, Wei-Ning and D'Avirro, Antony and Shi, Bowen and
               Gat, Itai and Fazel-Zarani, Maryam and Remez, Tal and Copet, Jade and
               Synnaeve, Gabriel and Hassid, Michael and Kreuk, Felix and Adi, Yossi and Dupoux, Emmanuel},
  booktitle = {Interspeech},
  year      = {2023}
}

@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}
}

License

CC-BY-NC-4.0 — non-commercial use only. See shangeth/expresso original_metadata/LICENSE.txt.