espnet/yodas3

YODAS v3

Paper

YODAS v3 is a large web-crawled dataset containing over 1.1 million hours of audio that were originally released under a CC-BY-3.0 license. The dataset contains audio in over 100 languages. YODAS v3 can be used for a variety of multi-modal tasks, including Automatic Speech Recognition, Text-to-Speech, and Audio Representation Learning. We crawl a distinct set of videos from the v1 and v2 versions of YODAS, to guarantee that there are no overlaps in the data.

For each audio file, we also release the following metadata:

• segment-level transcripts (if available)
• English translations (if available)
• language IDs
• audio length
• effective bandwith
• effective channel count

Approximately 75% of the data has transcripts. Around 60% of the non-English data has English translations. All transcripts and translations come with sentence-level timestamps. 95% of the transcripts also have word-level timestamps.

Language IDs are provided at the file-level, and may not be entirely accurate on individual sentences. The language IDs were obtained in 2 ways:

1. the language ID provided with the transcript, if available
2. the provided locale of the audio from the upload

While we originally reported the ID distribution from (2) in our paper, we found that the IDs from (1) are likely more accurate, if available. See the Analysis section below for more details. We therefore organize the upload directory by the language ID detected from (1), but include both IDs in the metadata. Any audio files without a transcript are put in the unk directory.

Check our blog post for information about potential use cases for YODAS v3, and how previous versions of the dataset were used.

Structure

data/
  <lang>/
    audio/0000.tar … NNNN.tar          # WebDataset shards, ≈8 GB each, members named <id>.webm
    metadata/0000.parquet … NNNN.parquet  # one parquet per shard, one row per member of the same-numbered tar

• 103 directories: 102 language codes (ISO-639-1 where one exists, otherwise fil, yue, zh, …) plus unk for audio with no caption track.
• 6,685 shards in total
• Each audio file has its own ID. metadata/NNNN.parquet has exactly the rows for the members of audio/NNNN.tar, joined on id, so a WebDataset pipeline can pair the two by shard number.

Metadata fields

field	dtype	values
id	string	32 lowercase hex characters; equals the tar member name without .webm; unique across the dataset
shard	string	4-digit shard number, e.g. "0042" — the tar this row's audio is in
transcript	string / null	Original-language captions as a JSON list of segments [{"start": <ms>, "duration": <ms>, "text": "…", "words": [{"t": <offset ms>, "w": "…"}, …]}, …]. words (per-word offsets relative to start) is present for ASR tracks (≈95 % of transcripts) and absent for uploader-provided subtitles. Null when the video has no caption track (all of unk, ≈2k others).
translation_en	string / null	English translation of a non-English original, same JSON format but segment-level only (no words). Null for English originals and where no translation exists (present for ≈1.07 M videos).
transcript_lang	string / null	Language of the original caption track = the <lang> directory; one of the 102 codes. Null in unk.
locale_lang	string / null	Language inferred from the uploader's locale (YODAS v1/v2 scheme), independent of the transcript; 146 distinct codes; null for ≈900 k videos.
length	float64	Audio duration in seconds (1 – 153,284; mean ≈ 952).
audio_bitrate	float64	Container bitrate in kbit/s (3.6 – 569, median ≈ 122); NaN for 1,461 videos.
max_freq_hz	float64 / null	Highest frequency with significant energy, from an FFT of a 10 s excerpt (35 – 24,000 Hz). Null for ≈283 k videos.
estimated_bandwidth_hz	float64 / null	Effective sample rate implied by max_freq_hz (urgent2025 method, −50 dB threshold): one of 8000, 16000, 22050, 24000, 32000, 44100. Use it to spot upsampled audio. Null where max_freq_hz is null.
channels	int32	Channel count in the file container: 1, 2, 4 or 6 (99.8 % are 2).
n_distinct_channels	int32 / null	Effective channel count after comparing channels on a 10 s excerpt (normalized RMS difference < 0.001 ⇒ identical): 1 – 5. channels = 2, n_distinct_channels = 1 (≈25 % of videos) means dual-mono. Null for 581 videos where the analysis failed.

Timestamps in transcript / translation_en are milliseconds from the start of the audio.

Analysis

Language coverage. YODAS v3 spans 100+ languages, and 34 of them have more than 1,000 hours, the rough threshold at which a language can support a strong standalone ASR model. As noted above, we measured language distribution using two different methods: locale and transcript. Below are example distribution figures for each measurement method:

Locale-based Language Distribution:

Transcript-based:

We originally reported the locale-based method in our paper.

Channel count. Over 70% of YODAS v3 contains at least 2 distinct channels. We counted a recording as multi-channel only when its channels were measurably different, so duplicated mono audio stored as stereo is labeled mono.

Effective sampling rate. Because much web audio is upsampled from lower-quality sources, we estimated each recording's effective sampling rate from its actual frequency content, rather than trusting the file header. Over 92% of the data reaches an effective rate of 32kHz or higher, and 67% reaches 44kHz.

Citation

We ask that you cite our paper if you use any of the resources we release under YODAS v3. If you re-release a large amount of data from YODAS v3 under a new dataset, benchmark, or repo, we also ask that put the below bibtex in any requested citations section. This makes it easier for us to get the resources needed for expanding the dataset in the future and keeping them free + open access.

@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},
}