--- license: cc-by-4.0 task_categories: - audio-to-audio - automatic-speech-recognition - text-to-speech - translation dataset_info: - config_name: preview features: - name: audio dtype: audio - name: lang dtype: string - name: id dtype: string - name: shard dtype: string - name: transcript dtype: string - name: translation_en dtype: string - name: transcript_lang dtype: string - name: locale_lang dtype: string - name: length dtype: float64 - name: audio_bitrate dtype: float64 - name: max_freq_hz dtype: float64 - name: estimated_bandwidth_hz dtype: float64 - name: channels dtype: int32 - name: n_distinct_channels dtype: int32 - name: preview_audio_seconds dtype: float64 configs: - config_name: preview default: true data_files: - split: train path: data/preview/*.parquet - config_name: ab_metadata data_files: - split: train path: data/ab/metadata/*.parquet - config_name: af_metadata data_files: - split: train path: data/af/metadata/*.parquet - config_name: am_metadata data_files: - split: train path: data/am/metadata/*.parquet - config_name: ar_metadata data_files: - split: train path: data/ar/metadata/*.parquet - config_name: as_metadata data_files: - split: train path: data/as/metadata/*.parquet - config_name: ay_metadata data_files: - split: train path: data/ay/metadata/*.parquet - config_name: az_metadata data_files: - split: train path: data/az/metadata/*.parquet - config_name: be_metadata data_files: - split: train path: data/be/metadata/*.parquet - config_name: bg_metadata data_files: - split: train path: data/bg/metadata/*.parquet - config_name: bn_metadata data_files: - split: train path: data/bn/metadata/*.parquet - config_name: bo_metadata data_files: - split: train path: data/bo/metadata/*.parquet - config_name: bs_metadata data_files: - split: train path: data/bs/metadata/*.parquet - config_name: ca_metadata data_files: - split: train path: data/ca/metadata/*.parquet - config_name: cr_metadata data_files: - split: train path: data/cr/metadata/*.parquet - config_name: cs_metadata data_files: - split: train path: data/cs/metadata/*.parquet - config_name: cy_metadata data_files: - split: train path: data/cy/metadata/*.parquet - config_name: da_metadata data_files: - split: train path: data/da/metadata/*.parquet - config_name: de_metadata data_files: - split: train path: data/de/metadata/*.parquet - config_name: el_metadata data_files: - split: train path: data/el/metadata/*.parquet - config_name: en_metadata data_files: - split: train path: data/en/metadata/*.parquet - config_name: eo_metadata data_files: - split: train path: data/eo/metadata/*.parquet - config_name: es_metadata data_files: - split: train path: data/es/metadata/*.parquet - config_name: et_metadata data_files: - split: train path: data/et/metadata/*.parquet - config_name: eu_metadata data_files: - split: train path: data/eu/metadata/*.parquet - config_name: fa_metadata data_files: - split: train path: data/fa/metadata/*.parquet - config_name: fi_metadata data_files: - split: train path: data/fi/metadata/*.parquet - config_name: fil_metadata data_files: - split: train path: data/fil/metadata/*.parquet - config_name: fo_metadata data_files: - split: train path: data/fo/metadata/*.parquet - config_name: fr_metadata data_files: - split: train path: data/fr/metadata/*.parquet - config_name: gl_metadata data_files: - split: train path: data/gl/metadata/*.parquet - config_name: gn_metadata data_files: - split: train path: data/gn/metadata/*.parquet - config_name: gu_metadata data_files: - split: train path: data/gu/metadata/*.parquet - config_name: ha_metadata data_files: - split: train path: data/ha/metadata/*.parquet - config_name: hi_metadata data_files: - split: train path: data/hi/metadata/*.parquet - config_name: hr_metadata data_files: - split: train path: data/hr/metadata/*.parquet - config_name: hu_metadata data_files: - split: train path: data/hu/metadata/*.parquet - config_name: hy_metadata data_files: - split: train path: data/hy/metadata/*.parquet - config_name: ia_metadata data_files: - split: train path: data/ia/metadata/*.parquet - config_name: id_metadata data_files: - split: train path: data/id/metadata/*.parquet - config_name: is_metadata data_files: - split: train path: data/is/metadata/*.parquet - config_name: it_metadata data_files: - split: train path: data/it/metadata/*.parquet - config_name: iw_metadata data_files: - split: train path: data/iw/metadata/*.parquet - config_name: ja_metadata data_files: - split: train path: data/ja/metadata/*.parquet - config_name: jv_metadata data_files: - split: train path: data/jv/metadata/*.parquet - config_name: ka_metadata data_files: - split: train path: data/ka/metadata/*.parquet - config_name: kk_metadata data_files: - split: train path: data/kk/metadata/*.parquet - config_name: km_metadata data_files: - split: train path: data/km/metadata/*.parquet - config_name: kn_metadata data_files: - split: train path: data/kn/metadata/*.parquet - config_name: ko_metadata data_files: - split: train path: data/ko/metadata/*.parquet - config_name: ku_metadata data_files: - split: train path: data/ku/metadata/*.parquet - config_name: ky_metadata data_files: - split: train path: data/ky/metadata/*.parquet - config_name: la_metadata data_files: - split: train path: data/la/metadata/*.parquet - config_name: lo_metadata data_files: - split: train path: data/lo/metadata/*.parquet - config_name: lt_metadata data_files: - split: train path: data/lt/metadata/*.parquet - config_name: lv_metadata data_files: - split: train path: data/lv/metadata/*.parquet - config_name: mk_metadata data_files: - split: train path: data/mk/metadata/*.parquet - config_name: ml_metadata data_files: - split: train path: data/ml/metadata/*.parquet - config_name: mn_metadata data_files: - split: train path: data/mn/metadata/*.parquet - config_name: mr_metadata data_files: - split: train path: data/mr/metadata/*.parquet - config_name: ms_metadata data_files: - split: train path: data/ms/metadata/*.parquet - config_name: mt_metadata data_files: - split: train path: data/mt/metadata/*.parquet - config_name: my_metadata data_files: - split: train path: data/my/metadata/*.parquet - config_name: ne_metadata data_files: - split: train path: data/ne/metadata/*.parquet - config_name: nl_metadata data_files: - split: train path: data/nl/metadata/*.parquet - config_name: no_metadata data_files: - split: train path: data/no/metadata/*.parquet - config_name: om_metadata data_files: - split: train path: data/om/metadata/*.parquet - config_name: or_metadata data_files: - split: train path: data/or/metadata/*.parquet - config_name: pa_metadata data_files: - split: train path: data/pa/metadata/*.parquet - config_name: pl_metadata data_files: - split: train path: data/pl/metadata/*.parquet - config_name: ps_metadata data_files: - split: train path: data/ps/metadata/*.parquet - config_name: pt_metadata data_files: - split: train path: data/pt/metadata/*.parquet - config_name: rm_metadata data_files: - split: train path: data/rm/metadata/*.parquet - config_name: ro_metadata data_files: - split: train path: data/ro/metadata/*.parquet - config_name: ru_metadata data_files: - split: train path: data/ru/metadata/*.parquet - config_name: rw_metadata data_files: - split: train path: data/rw/metadata/*.parquet - config_name: sa_metadata data_files: - split: train path: data/sa/metadata/*.parquet - config_name: sc_metadata data_files: - split: train path: data/sc/metadata/*.parquet - config_name: sd_metadata data_files: - split: train path: data/sd/metadata/*.parquet - config_name: si_metadata data_files: - split: train path: data/si/metadata/*.parquet - config_name: sk_metadata data_files: - split: train path: data/sk/metadata/*.parquet - config_name: sl_metadata data_files: - split: train path: data/sl/metadata/*.parquet - config_name: so_metadata data_files: - split: train path: data/so/metadata/*.parquet - config_name: sq_metadata data_files: - split: train path: data/sq/metadata/*.parquet - config_name: sr_metadata data_files: - split: train path: data/sr/metadata/*.parquet - config_name: su_metadata data_files: - split: train path: data/su/metadata/*.parquet - config_name: sv_metadata data_files: - split: train path: data/sv/metadata/*.parquet - config_name: sw_metadata data_files: - split: train path: data/sw/metadata/*.parquet - config_name: ta_metadata data_files: - split: train path: data/ta/metadata/*.parquet - config_name: te_metadata data_files: - split: train path: data/te/metadata/*.parquet - config_name: tg_metadata data_files: - split: train path: data/tg/metadata/*.parquet - config_name: th_metadata data_files: - split: train path: data/th/metadata/*.parquet - config_name: ti_metadata data_files: - split: train path: data/ti/metadata/*.parquet - config_name: tk_metadata data_files: - split: train path: data/tk/metadata/*.parquet - config_name: tr_metadata data_files: - split: train path: data/tr/metadata/*.parquet - config_name: uk_metadata data_files: - split: train path: data/uk/metadata/*.parquet - config_name: unk_metadata data_files: - split: train path: data/unk/metadata/*.parquet - config_name: ur_metadata data_files: - split: train path: data/ur/metadata/*.parquet - config_name: uz_metadata data_files: - split: train path: data/uz/metadata/*.parquet - config_name: vi_metadata data_files: - split: train path: data/vi/metadata/*.parquet - config_name: wo_metadata data_files: - split: train path: data/wo/metadata/*.parquet - config_name: yue_metadata data_files: - split: train path: data/yue/metadata/*.parquet - config_name: zh_metadata data_files: - split: train path: data/zh/metadata/*.parquet - config_name: zu_metadata data_files: - split: train path: data/zu/metadata/*.parquet --- ## YODAS v3 [Paper](https://www.isca-archive.org/interspeech_2026/chen26d_interspeech.html) 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](https://huggingface.co/datasets/espnet/yodas) and [v2](https://huggingface.co/datasets/espnet/yodas2) 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](https://huggingface.co/blog/espnet/yodasv3) for information about potential use cases for YODAS v3, and how previous versions of the dataset were used. ## Downloading ### HF Transfer **One language only**. ``` hf download espnet/yodas3 --repo-type dataset --include "data/hi/*" --local-dir yodas3 ``` **Multiple languages**. ``` hf download espnet/yodas3 --repo-type dataset \ --include "data/hi/*" "data/ta/*" "data/bn/*" --local-dir yodas3 ``` **Just the metadata for all languages (no audio)** ``` hf download espnet/yodas3 --repo-type dataset --include "data/*/metadata/*" --local-dir yodas3 ``` **Specific shards** ``` hf download espnet/yodas3 --repo-type dataset \ --include "data/en/audio/00[0-4]?.tar" "data/en/metadata/00[0-4]?.parquet" --local-dir yodas3 ``` ### load_dataset ``` """Join espnet/yodas3 audio (WebDataset tars) with transcripts (metadata parquets), streaming. python join_example.py hi # stream Hindi, print the first few joined samples """ import io, json, sys import av import numpy as np from datasets import load_dataset lang = sys.argv[1] if len(sys.argv) > 1 else "hi" N = int(sys.argv[2]) if len(sys.argv) > 2 else 3 # 1) metadata: one row per video, keyed by the encrypted id (= tar member name) meta = {r["id"]: r for r in load_dataset("espnet/yodas3", f"{lang}_metadata", split="train")} # 2) audio: stream the tars; each sample = {"__key__": id, "__url__": shard url, "webm": bytes} audio = load_dataset("webdataset", data_files=f"hf://datasets/espnet/yodas3/data/{lang}/audio/*.tar", split="train", streaming=True) def decode_webm(b: bytes, sr: int = 16000) -> np.ndarray: """Opus-in-WebM bytes -> float32 mono waveform at `sr` Hz, in-process via PyAV (pip install av).""" with av.open(io.BytesIO(b)) as c: rs = av.AudioResampler(format="flt", layout="mono", rate=sr) chunks = [f.to_ndarray().reshape(-1) for frame in c.decode(c.streams.audio[0]) for f in rs.resample(frame)] chunks += [f.to_ndarray().reshape(-1) for f in rs.resample(None)] # flush the resampler return np.concatenate(chunks) for i, s in enumerate(audio): m = meta[s["__key__"]] # 3) the join segments = json.loads(m["transcript"]) if m["transcript"] else [] wav = decode_webm(s["webm"]) print(f"{s['__key__']} shard={m['shard']} {m['length']:.0f}s decoded={len(wav)/16000:.1f}s " f"{len(segments)} segments first={segments[0]['text'][:60]!r}" if segments else f"{s['__key__']} shard={m['shard']} {m['length']:.0f}s decoded={len(wav)/16000:.1f}s no transcript") if i + 1 >= N: break ``` ### load_dataset (pure streaming, no parquet download) ``` from datasets import load_dataset, Features, Value, List B = "hf://datasets/espnet/yodas3/data" WORD = {"t": Value("int64"), "w": Value("string")} SEG = {"start": Value("int64"), "duration": Value("int64"), "text": Value("string"), "words": List(WORD)} TSEG = {"start": Value("int64"), "duration": Value("int64"), "text": Value("string")} FEATURES = Features({ "__key__": Value("string"), "__url__": Value("string"), "webm": Value("binary"), "json": {"id": Value("string"), "shard": Value("string"), "transcript": List(SEG), "translation_en": List(TSEG), "transcript_lang": Value("string"), "locale_lang": Value("string"), "length": Value("float64"), "audio_bitrate": Value("float64"), "max_freq_hz": Value("float64"), "estimated_bandwidth_hz": Value("float64"), "channels": Value("int64"), "n_distinct_channels": Value("int64")}}) # one language ds = load_dataset("webdataset", data_files=f"{B}/hi/audio/*.tar", split="train", streaming=True, features=FEATURES) # several languages ds = load_dataset("webdataset", data_files=[f"{B}/hi/audio/*.tar", f"{B}/ta/audio/*.tar", f"{B}/bn/audio/*.tar"], split="train", streaming=True, features=FEATURES) # high bandwith only via .filter() ds = load_dataset("webdataset", data_files=f"{B}/*/audio/*.tar", split="train", streaming=True, features=FEATURES).filter(lambda s: (s["json"]["estimated_bandwidth_hz"] or 0) > 32000) # all ds = load_dataset("webdataset", data_files=f"{B}/*/audio/*.tar", split="train", streaming=True, features=FEATURES) ``` # WebDataset ``` import webdataset as wds U = "https://huggingface.co/datasets/espnet/yodas3/resolve/main/data" # one language ds = wds.WebDataset(f"{U}/hi/audio/{{0000..0417}}.tar", shardshuffle=False).decode() # several languages ds = wds.WebDataset(f"{U}/hi/audio/{{0000..0417}}.tar::{U}/ta/audio/{{0000..0020}}.tar::{U}/bn/audio/{{0000..0040}}.tar", shardshuffle=False).decode() # filter for high bandwith ds = wds.WebDataset(f"{U}/hi/audio/{{0000..0417}}.tar", shardshuffle=False).decode().select(lambda s: (s["json"]["estimated_bandwidth_hz"] or 0) > 32000) # all from huggingface_hub import HfFileSystem ds = wds.WebDataset([f"{U}/{p.split('/data/',1)[1]}" for p in HfFileSystem().glob("datasets/espnet/yodas3/data/*/audio/*.tar")], shardshuffle=False).decode() ``` ## Structure ``` data/ / audio/0000.tar … NNNN.tar # WebDataset shards, ≈8 GB each, members named .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": , "duration": , "text": "…", "words": [{"t": , "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 `` 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: ![image](https://cdn-uploads.huggingface.co/production/uploads/630438615c70c21d0eae6613/v0miBFP7mStIcR68LK9Qe.png) Transcript-based: ![image](https://cdn-uploads.huggingface.co/production/uploads/630438615c70c21d0eae6613/OzJZukwyp8HqX81oJo4eu.png) 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. ![image](https://cdn-uploads.huggingface.co/production/uploads/630438615c70c21d0eae6613/nrDBvKQK99nDJtSDHEHVq.png) **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}, } ```