Datasets:

shangeth
/

expresso-mimi-codes-tagged

id stringlengths 14 48	text stringlengths 2 2.51k	speaker_id stringclasses 4 values	codes listlengths 32 32	n_frames int32 4 2.32k	k_codebooks int32 32 32
ex01_default_longform_00001	"Yellowstone National Park is an American national park located in the western United States, largel(...TRUNCATED)	1	[[1049,958,1345,826,129,347,1435,1651,759,638,760,1517,682,331,574,1023,306,1338,1338,1327,1327,1196(...TRUNCATED)	2,062	32
ex01_narration_longform_00001	"<narration> Cherry releases the grip around her brother, steadying her trembling feet onto the hot,(...TRUNCATED)	1	[[1049,127,1798,1041,902,1460,1470,862,473,96,1472,735,1451,1251,1483,2045,109,109,1668,1295,1765,17(...TRUNCATED)	2,124	32
ex02_default_longform_00001	"Stefani Joanne Angelina Germanotta, born March twenty-eighth, nineteen eighty six, known profession(...TRUNCATED)	2	[[1049,127,738,201,529,784,1383,584,336,908,406,99,1611,1611,300,342,908,1939,348,1791,1071,908,1764(...TRUNCATED)	1,760	32
ex02_narration_longform_00001	"<narration> I was born at Woodstock. At two minutes past three on the morning of Monday, August eig(...TRUNCATED)	2	[[1049,958,1972,1494,982,1546,1001,382,794,1908,680,1184,67,1948,1250,997,622,561,561,833,644,447,61(...TRUNCATED)	2,324	32
ex03_default_longform_00001	"Three astronauts successfully reached the International Space Station this morning, where their six(...TRUNCATED)	3	[[1049,958,1641,357,357,1384,1421,2027,1500,132,720,837,1185,1040,1040,252,1571,1424,682,307,722,147(...TRUNCATED)	1,674	32
ex03_narration_longform_00001	"<narration> The view from the second-floor terrace was panoramic, and breathtaking. Justine Nolan, (...TRUNCATED)	3	[[1049,958,246,419,655,1077,1077,1164,1549,440,1047,1591,1134,252,1580,808,1416,1525,1344,707,847,18(...TRUNCATED)	1,873	32
ex04_default_longform_00001	"Toxic chemicals are accumulating in marine creatures in Earth’s deepest oceanic trenches, the fir(...TRUNCATED)	4	[[1049,958,1081,447,984,895,28,1431,1708,1708,996,1624,1405,28,785,1077,1033,1939,501,1891,1932,1479(...TRUNCATED)	1,688	32
ex04_narration_longform_00001	"<narration> Nineteen forty-eight, Piedmont, Northern Italy. The russet bloom on the vineyards ahead(...TRUNCATED)	4	[[1049,958,675,104,1192,1412,1207,1113,957,1416,1525,1192,357,471,1966,560,1877,1039,857,1987,729,39(...TRUNCATED)	1,734	32
ex01_confused_00001	<confused> Why are you beating up my jukebox?	1	[[1049,958,1962,1187,1187,997,927,927,75,1384,65,557,1874,162,287,1623,1932,1437,886,1264,629,1281,9(...TRUNCATED)	35	32
ex01_confused_00002	<confused> I have to stop you.	1	[[1049,958,161,1972,999,1838,1488,376,1335,1849,1339,2030,1323,1967,1245,1161,833,644,1560,254,786,7(...TRUNCATED)	29	32

End of preview. Expand in Data Studio

Expresso — Tagged Mimi Codes (k=32)

Training-ready derivative of shangeth/expresso-mimi-codes, built specifically for style-conditioned TTS fine-tuning.

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

What this differs from `expresso-mimi-codes`

This dataset:

Merges read + conversational configs into a single flat dataset per split (matches the canonical schema other *-mimi-codes datasets use).
Drops 5 styles whose ASR transcripts are not reliably aligned to the audio (the speaker is doing voice acting / non-verbal sounds): animal, animaldir, child, childdir, nonverbal.
Prepends a style tag to the text for 19 tagged styles. default rows are left untagged. The model learns "no tag = default voice, <style> = stylized delivery".

Schema

Column	Type	Notes
`id`	string	unchanged
`text`	string	tagged: `<style> {text}` for the 19 tags; bare for `default`
`speaker_id`	string	cast from int (1–4)
`codes`	`int16[32][n_frames]`	all 32 Mimi codebooks @ 12.5 fps; slice `codes[:k]` for fewer
`n_frames`	int32
`k_codebooks`	int32	32

The 19 tags

EMOTIONAL    <happy> <sad> <angry> <fearful> <disgusted>
             <awe> <desire> <calm> <sympathetic>
DELIVERY     <laughing> <enunciated> <whisper> <fast> <projected>
PERFORMANCE  <confused> <sarcastic> <narration>
STATE        <bored> <sleepy>

Plus untagged default (~250 min / 4.2 h) as the no-tag baseline.

Splits

split	rows	hours
train	~26k	~37
dev	~1.1k	~1.4
test	~1.1k	~1.4

Usage

from datasets import load_dataset
import torch

ds = load_dataset("shangeth/expresso-mimi-codes-tagged", split="train")
ex = ds[0]
print(ex["text"])              # e.g. "<happy> Hello, how are you?"
print(ex["speaker_id"])        # "1" .. "4"
codes = torch.tensor(ex["codes"], dtype=torch.long)  # [32, n_frames]

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

Reproducing this dataset

python expresso_tagged.py \\
  --src_repo shangeth/expresso-mimi-codes \\
  --dst_repo shangeth/expresso-mimi-codes-tagged --private

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}
}