You need to agree to share your Hugging Face information to access this dataset
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
By requesting access you agree to share your name, company and email with Trelis, and to receive an email from Trelis confirming access.
Log in or Sign Up to review the conditions and access this dataset content.
YODAS3 tokenized speech-to-speech pre-training dataset
For speech separation and audio annotation services, contact voice@trelis.com.
A filtered and tokenized pre-training dataset for dual-track speech-to-speech models. It holds long spans of English speech with one or two speakers, chosen to hold little or no music. Drawn from YODAS3 (English), referenced by shard and member.
The dataset ships audio tokens in place of raw audio, which keeps it compact. It carries rich machine annotations: speaker-attributed transcripts with word timings, including overlapping speech; vocal bursts such as laughter; per-utterance emotion labels; and, where two speakers are present, each speaker on their own token track.
Separated speaker audio is included. On 856 of the 3630 spans two people speak, and each speaker has their own audio token track alongside the mixed audio, with that speaker's word timings.
- Size: 3630 spans, 408.25 hours.
- Audio tokens: Mimi codes for each span, 8 codebooks at 12.5 Hz from 24 kHz mono audio. Load with
numpy.load("<folder>/tokens.npz")[key]. 856 of the 3630 spans also have one array per speaker,<key>-track1and<key>-track2, alongside the mixed array;separation.statusisokon those rows. - Raw audio: not included. Each row links to its source and gives the span's sample bounds in it,
source_start_sampleandsource_end_sample. A source that has changed since, for example through inserted advertising, no longer matches the recorded hashes. - Language: English.
- Transcripts: 3623 of the 3630 spans have a speaker-attributed machine transcript with word timings, and 512 have per-utterance emotion labels. The emotion labels are mostly
happyandneutral; other classes are rare and less reliable. - Labels: every label is machine-generated and has not been corrected by hand.
Columns
One row per span, in <folder>/train.parquet, <folder>/validation.parquet or <folder>/test.parquet by its split. A nested field a row does not have is null.
| Column | Content |
|---|---|
recording_id, source |
Source recording and where to fetch it, with content hashes |
episode_id, span_index, start_s, end_s |
The episode or recording, the span's position among its rows, and its time in it; rows are in episode order |
start_sample, end_sample, audio, decoder |
Span position in the decoded piece that audio describes; for podcasts a piece is one slice of the episode |
source_start_sample, source_end_sample, seconds |
Span position in the source recording at its native sample rate, and duration. Cut the span from the source with these |
music_fraction |
Share of the span with music |
dataset |
The sets of this family the span belongs to |
diarization |
Speaker turns |
stats |
Conversation statistics: speakers, speech share, turn and overlap rates, gaps |
tokens |
Record of the span's mixed-audio token array; tokens.key names the array in the folder's tokens.npz |
tracks |
Two speaker tracks, each with its token record and its speaker's word timings. With two speakers each track is one speaker; with one speaker track 2 is empty |
separation |
separation.status is ok where the two tracks are separate per-speaker arrays, <key>-track1 and <key>-track2 |
vocal_bursts |
Laughter and other non-speech vocal sounds with labels and sample positions on the start_sample clock |
conversational, emotional |
Subset flags |
split, split_group |
train, validation or test, and the unit kept whole in one split: a podcast feed or a YODAS3 recording has the same split in every dataset. Score a model on the splits of the revision it trained on |
transcription_status, transcript, segments, speakers |
Speaker-attributed machine transcript; the status is ok where one exists |
words |
Word timings with speaker |
utterances, emotion_status, emotion_seconds |
Per-utterance emotion labels and seconds per emotion |
delivery |
Speaking-rate and turn-taking measures |
selection |
Descriptive tags such as balanced or high_overlap |
Each folder's report.json holds its span and hour counts and an internal_reference block of service revisions, kept for our own traceability.
The transcript, word timing, emotion and delivery columns are empty on spans without a transcript.
Usage
Rows come in train, validation and test splits, one Parquet file per split in each folder, so datasets.load_dataset("Trelis/duplex-pre-training-yodas3-mimi", split="validation") loads the rows of a split. Filter rows with ordinary Python. Useful fields: stats.speakers (1 or 2), separation.status (ok where each speaker has their own track), conversational, emotional, selection.tags (for example high_overlap, rapid_turns, balanced), music_fraction and split.
Decode tokens to audio with the Mimi codec to listen to a span, mixed and per speaker:
import numpy as np
import pyarrow.parquet as pq
import soundfile as sf
import torch
from huggingface_hub import snapshot_download
from transformers import MimiModel
repo = "Trelis/duplex-pre-training-yodas3-mimi"
folder = "r3-yodas-2" # one folder of the dataset; manifest.json lists them all under "runs"
root = snapshot_download(repo, repo_type="dataset", allow_patterns=[f"{folder}/*"])
rows = pq.read_table(f"{root}/{folder}/train.parquet").to_pylist()
tokens = np.load(f"{root}/{folder}/tokens.npz")
# Two-speaker spans with each speaker on their own track; add r["conversational"] for lively turn-taking.
separated = [r for r in rows if r["stats"]["speakers"] == 2 and r["separation"]["status"] == "ok"]
row = separated[0]
mimi = MimiModel.from_pretrained("kyutai/mimi").eval()
def decode(key, seconds=20):
codes = torch.from_numpy(tokens[key][:, : int(seconds * 12.5)].astype(np.int64))[None]
with torch.no_grad():
return mimi.decode(codes).audio_values[0, 0].numpy()
sf.write("mixed.wav", decode(row["tokens"]["key"]), 24000)
for number, track in enumerate(row["tracks"], 1):
if track["status"] == "ok":
sf.write(f"speaker{number}.wav", decode(track["tokens"]["key"]), 24000)
print(number, " ".join(w["word"] for w in track["words"][:12]))
Licence
CC BY 3.0, following YODAS3. If you use this dataset, cite YODAS3:
@inproceedings{chen26d_interspeech,
title = {{YODAS v3: Over 1 Million Hours of High-Bandwidth, Stereophonic, Multilingual Speech}},
author = {William Chen and Shinnosuke Takamichi and Sayaka Shiota and Satoru Fukayama and Samuele Cornell and Shinji Watanabe},
year = {2026},
booktitle = {{Interspeech 2026}},
pages = {5367--5372},
doi = {10.21437/Interspeech.2026-386},
issn = {2958-1796},
}
- Downloads last month
- 37