Datasets:

shangeth
/

vctk-mimi-codes

license: cc-by-4.0
language:
  - en
task_categories:
  - text-to-speech
tags:
  - mimi
  - neural-codec
  - speech-synthesis
  - vctk
  - audio-tokens
  - accents
pretty_name: VCTK Mimi Codes (mic1)
size_categories:
  - 10K<n<100K

VCTK — Mimi Codes (mic1)

Pre-extracted Kyutai Mimi tokens for the VCTK Corpus — 109 speakers across 11 British, Scottish, and American accents. ~44h of read speech.

Only mic1 recordings are included. Each utterance was recorded with two microphones; mic1 (close microphone) gives a cleaner signal. Mic2 duplicates are excluded. Utterance IDs end in _mic1 (e.g. p225_001_mic1).

Schema

Column	Type	Notes
`id`	string	e.g. `p225_001_mic1`
`text`	string	read sentence, mixed-case with punctuation
`speaker_id`	int32	numeric speaker ID (225 for p225)
`accent`	string	e.g. `English`, `Scottish`, `American`
`codes`	`int16[k=8][n_frames]`	Mimi codebook indices @ 12.5 fps
`n_frames`	int32
`k_codebooks`	int32	8

Extraction details

Source: sanchit-gandhi/vctk
Codec: kyutai/mimi @ 24 kHz, 12.5 fps
Resampling: 48 kHz → 24 kHz
Filter: file column stem must end with _mic1

Usage

from datasets import load_dataset
import torch

ds = load_dataset("shangeth/vctk-mimi-codes", split="train")
ex = ds[0]
codes = torch.tensor(ex["codes"], dtype=torch.long)  # [8, n_frames]
print(ex["id"], ex["accent"], "→", ex["text"])

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{veaux2017cstr,
  title     = {CSTR VCTK Corpus: English Multi-speaker Corpus for CSTR Voice Cloning Toolkit},
  author    = {Veaux, Christophe and Yamagishi, Junichi and MacDonald, Kirsten},
  year      = {2017}
}

License

CC-BY-4.0.