destitech/custom-yodas-multi

Custom YODAS Multi

A filtered, sentence-level, feature-only multilingual text-to-speech dataset built from YODAS2: 28 languages, 869,096 sentences, 1,290 hours, 70,855 source videos.

It contains no audio. Each sentence comes with a log-mel spectrogram, an F0 track, its transcript, per-character durations, a speaker-timbre embedding, emotion scores and recording-quality scores. Turning the mels back into sound requires a vocoder trained for this exact mel configuration.

Contents

Path	What
metadata/<lang>.parquet	one row per sentence (this is what the dataset viewer shows)
features/<lang>/<lang>-NNNNN.tar	~1 GiB shards holding <segment_id>.mel.npy and <segment_id>.world.npz
stats.json	per-language counts

Languages

Code	Language	Sentences	Hours	Videos	ASR used for the transcript
de	German	86,628	136.2	6,102	Parakeet TDT v3
tr	Turkish	86,166	118.5	7,994	Whisper large-v3-turbo
fr	French	66,088	106.3	6,795	Parakeet TDT v3
ru	Russian	74,922	104.2	5,804	Parakeet TDT v3
en	English	61,288	100.5	5,098	Parakeet TDT v3
es	Spanish	70,765	99.9	5,217	Parakeet TDT v3
it	Italian	67,350	96.1	5,572	Parakeet TDT v3
pt	Portuguese	64,943	95.9	5,846	Parakeet TDT v3
pl	Polish	61,739	88.5	4,124	Parakeet TDT v3
uk	Ukrainian	36,341	69.8	5,007	Parakeet TDT v3
nl	Dutch	57,645	69.5	5,961	Parakeet TDT v3
ar	Arabic	22,321	50.2	1,840	NVIDIA FastConformer Arabic
hu	Hungarian	30,245	36.0	373	Parakeet TDT v3
fi	Finnish	25,342	35.0	846	Parakeet TDT v3
fa	Persian	9,272	16.3	231	Whisper large-v3-turbo
no	Norwegian	9,702	15.1	620	Whisper large-v3-turbo
sv	Swedish	11,389	14.5	684	Parakeet TDT v3
iw	Hebrew	10,742	13.6	730	Whisper large-v3-turbo
cs	Czech	6,169	8.5	762	Parakeet TDT v3
el	Greek	3,317	5.2	403	Parakeet TDT v3
ca	Catalan	2,207	3.5	270	Whisper large-v3-turbo
ro	Romanian	1,638	2.5	209	Parakeet TDT v3
bg	Bulgarian	892	1.5	78	Parakeet TDT v3
da	Danish	805	1.2	126	Parakeet TDT v3
sk	Slovak	485	0.7	53	Parakeet TDT v3
hr	Croatian	349	0.5	50	Parakeet TDT v3
sl	Slovenian	336	0.4	57	Parakeet TDT v3
be	Belarusian	10	0.0	3	Whisper large-v3-turbo
	Total	869,096	1,290.1	70,855

Folder and config names keep the YODAS codes (iw = Hebrew). Six languages hold less than 2 hours (be, sl, hr, da, sk, bg); they are included as-is.

Metadata columns

Column	Type	Meaning
segment_id	string	<youtube_id>__<utterance>_<sentence>; also the file stem inside the shards
manifest_index	int	row position in the original (shuffled) training manifest
language	string	YODAS language code
video_id	string	YouTube id of the source video (attribution)
utterance_id	string	the VAD utterance the sentence was cut from
is_first_chunk	bool	first sentence of its utterance
text	string	normalised transcript; may contain event tokens such as <laugh>, <breath>, <sigh>
n_frames, duration_s	int, float	mel length (10 ms frames)
char_frames	list[int]	frames per text unit (one unit per character, event tokens count as one unit)
timbre	list[float] ×128	Orange/Speaker-wavLM-tbr embedding
emotion	list[float] ×4	LAION Empathic-Insight scores: expressiveness (monotone ↔ expressive), concentration, impatience/irritability, hope/enthusiasm/optimism
dnsmos_ovrl_mean, dnsmos_ovrl_min	float	DNSMOS P.835 OVRL over 1–3 segments of the sentence
snr_db	float	estimated signal-to-noise ratio
shard	string	tar file holding the features

Feature files

• <segment_id>.mel.npy — float16, shape (100, T). Natural-log mel, 24 kHz audio, n_fft 1024, hop 240 (10 ms), 100 mel bins, loudness-normalised source (−23 LUFS).
• <segment_id>.world.npz — key f0: float64, one value per 10 ms frame (length T or T+1), Hz, 0 on unvoiced frames (WORLD harvest, 100–1000 Hz, median filter 3, 10 ms frame period).

Usage

Metadata only:

from datasets import load_dataset
ds = load_dataset("destitech/custom-yodas-multi", "fr", split="train")   # or no config name for all languages

Features of one shard:

import io, tarfile, numpy as np
from huggingface_hub import hf_hub_download

row = ds[0]
tar = tarfile.open(hf_hub_download("destitech/custom-yodas-multi", row["shard"], repo_type="dataset"))
mel = np.load(io.BytesIO(tar.extractfile(row["segment_id"] + ".mel.npy").read()))             # (100, T) float16
f0 = np.load(io.BytesIO(tar.extractfile(row["segment_id"] + ".world.npz").read()))["f0"]      # (T or T+1,)

The shards follow the WebDataset naming convention (key.mel.npy, key.world.npz), sorted by key, so they can also be streamed with webdataset.

How it was built

Up to 2,000 hours of raw YODAS2 audio per language (every language with at least 20 raw hours, Asian languages excluded) went through this pipeline; roughly 5–10 % survived.

1. Loudness normalisation to −23 LUFS at 24 kHz; VAD split into utterances of 1–28 s.
2. Screening: SNR ≥ 12 dB, clipping ratio ≤ 0.1 %.
3. Sound-event detection (BEATs) to tag laughter, breaths, sighs… and to reject music and singing.
4. Transcription: NVIDIA Parakeet TDT v3 where it supports the language, Whisper large-v3-turbo otherwise (with a hallucination filter), NVIDIA FastConformer for Arabic. Text normalisation of numbers and abbreviations where a deterministic normaliser exists.
5. Forced alignment (MMS) to character level, then sentence splitting on pauses (0.6–20 s sentences).
6. Filtering: spoken-language identification must match the YODAS label (confidence ≥ 0.8), LAION overall quality ≥ 2.5 and background quality ≥ 3.0, SNR ≥ 15 dB, DNSMOS OVRL mean ≥ 2.7 and minimum ≥ 2.4 over the sentence's segments, alignment score and speaking-rate sanity checks.
7. Timbre embedding, mel and F0 extraction.

Limitations

• Transcripts are machine-generated and not human-verified. Errors remain, more so in the languages transcribed with Whisper and in the smallest languages.
• Language labels are noisy at the source; a language-ID filter removed most mismatches, not all.
• Uneven coverage: from 136 hours down to a few minutes per language.
• No speaker labels. The timbre embedding is the only speaker information; one video may contain several speakers.
• Spontaneous YouTube speech: hesitations, informal register, variable recording conditions.

Personal data and intended use

The source is public YouTube speech from real people, and timbre embeddings characterise their voices: this is personal data. The dataset is meant for research and development of speech synthesis. Do not use it to identify or re-identify speakers, to impersonate a real person, or to clone a specific person's voice without their consent.

Removal requests. If you are the author of, or appear in, a source video and want it removed, open a discussion in the Community tab of this repository with the video id; the corresponding rows and features will be deleted.

Licence and attribution

Released under CC-BY-4.0.

This is a derivative of YODAS2 (CC-BY-3.0), which collects YouTube videos published by their uploaders under a Creative Commons Attribution licence. Each row keeps the YouTube video_id so the original author can be credited (https://www.youtube.com/watch?v=<video_id>). The Creative Commons status of each video is the uploader's own declaration and is not verified here.

If you use this dataset, please credit YODAS as well:

@inproceedings{li2023yodas,
  title={Yodas: Youtube-Oriented Dataset for Audio and Speech},
  author={Li, Xinjian and Takamichi, Shinnosuke and Saeki, Takaaki and Chen, William and Shiota, Sayaka and Watanabe, Shinji},
  booktitle={2023 IEEE Automatic Speech Recognition and Understanding Workshop (ASRU)},
  year={2023}
}