psdn-ai/CROWS-multilingual

CROWS Multilingual: CRowdsourced Open World Speech in 16 Languages

Crowdsourced Open World Speech (CROWS) Multilingual is a speech corpus collected by Poseidon AI, Inc. It contains 821,361 scripted read-speech recordings from 46,298 speaker profiles across 16 languages. Every recording comes with its reading prompt, an automatic transcript, an English translation, and per-recording quality and speaker-consistency scores. Recordings from contributors flagged as likely synthetic-speech submitters are separated into their own partitions, so you can use the data for multilingual ASR and for synthetic-speech (deepfake) detection research.

Need a custom speech dataset? Talk to Poseidon →

Corpus	821,361 recordings · ~939 GB of original WAV audio
Languages	16: Arabic, German, English, Spanish, French, Hindi, Indonesian, Japanese, Korean, Marathi, Portuguese, Russian, Turkish, Urdu, Vietnamese, Mandarin Chinese
Speakers	46,298 pseudonymous speaker profiles. 12,719 profiles recorded in more than one language.
Speech type	Scripted, single-speaker read speech
Prompts	51,885 unique reading prompts. Each non-English prompt includes an English back-translation and a link to its English source prompt.
Transcripts	Automatic Whisper-large transcript plus an English translation for every recording
Quality signals	Per-recording WER, CER, semantic similarity to the prompt, and speaker-consistency score
Synthetic-risk partitions	8,897 recordings from contributors flagged as higher risk, kept separate from the main data
Audio	Original WAV files, byte-for-byte (no resampling, normalization, or transcoding), available embedded in Parquet or as TAR archives
Integrity	Every recording decode-checked; SHA-256 checksum for every WAV file and every published artifact (see release_manifest.json)

Main and Synthetic Risk Partitions

Main data, archives, checksums, and root metadata.csv contain 812,464 recordings. The two disjoint risk partitions are in synthetic_risk:

• data_from_users_with_breadth_gt_6: 2,260 recordings, relocated from suspect_data, selected by breadth_gt_6.csv.
• data_from_users_with_pred_synthetic_proportion_gt_0_3: 6,637 recordings from the 496 profiles listed in predicted_synthetic_proportion_users_gt_0_3.csv, moved out of main data.

Each partition contains a global metadata CSV, per-language CSVs, Parquet audio, TAR WAVs and JSON sidecars, and original checksums. All metadata uses the same twelve columns; WAV filenames remain <file_id>.wav. Selection criteria flag risk rather than establish synthetic ground truth.

Use breadth_gt_6_all / breadth_gt_6_<language> and pred_synthetic_gt_0_3_all / pred_synthetic_gt_0_3_<language> configurations, each with the full split. Default all and language configurations contain main data only.

from datasets import load_dataset
main = load_dataset("psdn-ai/CROWS-multilingual", "en", split="full", streaming=True)
breadth = load_dataset("psdn-ai/CROWS-multilingual", "breadth_gt_6_en", split="full", streaming=True)
proportion = load_dataset("psdn-ai/CROWS-multilingual", "pred_synthetic_gt_0_3_en", split="full", streaming=True)

Files and viewer

• metadata.csv is the main recording index, with exactly twelve columns.
• archives/<language>/full-XXXXX.tar contains original WAVs named exactly <file_id>.wav at the archive root and <file_id>.json metadata sidecars.
• data/<language>/full-XXXXX.parquet embeds the same original WAV bytes with an audio column first, followed by the twelve metadata columns for playback and filtering.
• checksums/<language>/full-XXXXX.jsonl records the SHA-256, byte count, sample rate, channel count, and frame count of each WAV.
• release_manifest.json records final counts and artifact checksums.

The TAR and Parquet representations contain the same recordings. Download only the representation you need. No resampling, normalization, or transcoding is applied when packaging the original WAV files.

Use the dataset viewer's configuration dropdown to select a language, or all for the combined release. Metadata columns expose language, user ID, script ID, and file ID alongside playable audio. The CSV is a standalone index; it is not an AudioFolder loader for the contents of the TAR archives.

Script prompts

Browse prompts by language · Open the scripts viewer · Loading examples

Readable pages show every prompt with its back translation and linked English content. Language-specific JSONL downloads make it easy to parse one prompt per line.

scripts/scripts.csv provides 51,885 unique prompts for every prompt_id in the audio metadata. Join audio prompt_id to prompt prompt_id. The scripts viewer configuration uses an identical Parquet copy and shows prompt content, its back translation, and the linked English prompt ID and content. See scripts/README.md for the exact database field mapping and handling of missing links.

Metadata schema

Release v1.8 refreshes semantic_score from files.semantic_score_std_v3 as exported on October 15, 2026. All 821,361 records matched; 811,154 scores changed and none are missing. In every audio viewer configuration, audio is the first column, followed by the twelve metadata fields. CSVs and TAR JSON sidecars retain their twelve-field metadata schema. See refresh report.

In release v1.5, user_id was renamed to profile_id with values unchanged, and expected_language was removed from audio metadata.

All main and synthetic risk audio CSVs, Parquet rows, and TAR JSON sidecars use the same twelve metadata fields. Parquet additionally contains the audio feature. script_id has been renamed to prompt_id; its values are unchanged.

Column	Parquet type	Source / meaning
file_name	string	Bare WAV filename, <file_id>.wav
profile_id	string	Preserved user identifier from the source filename
language	string	Preserved language code from the source filename
prompt_id	string	Former script_id; joins to the shared prompt table's prompt_id
file_id	string	Recording identifier; joins to files.id and embedding-table file_id
anchor_cos_similarity	float64, nullable	speechbrain_spkrec_ecapa_voxceleb_embeddings_v2.anchor_similarity
duration	float64	files.duration, copied as stored
whisper_large_hypothesis_transcript	string	files.spoken_native
english_translated_hypothesis_transcript	string	files.spoken_english
wer	float64	files.wer_std_v3
cer	float64	files.cer_std_v3
semantic_score	float64	files.semantic_score_std_v3

file_name,profile_id,language,prompt_id,file_id,anchor_cos_similarity,duration,whisper_large_hypothesis_transcript,english_translated_hypothesis_transcript,wer,cer,semantic_score

Fields were joined by the exact recording IDs selected by passing_files.csv. All 821,361 recordings match files; 1,592 have no embedding-table row, so their anchor_cos_similarity is null. All other requested fields are populated.

Numeric values retain the source scale and units. No scores are recomputed. CSV represents missing values as empty cells; Parquet retains native nullable numeric and text types. Transcript text is preserved, including commas, quotes, Unicode, and line breaks. Parse CSV with a CSV reader rather than splitting on commas or physical lines.

See metadata_enrichment_report.json for coverage and source mappings. WAV filenames and original audio bytes remain unchanged across all three partitions.

Language counts

Counts below include main and both synthetic risk partitions. Their individual counts are recorded in release_manifest.json and the synthetic risk folder indexes.

Code	Language	Recordings
ar	Arabic	14,236
de	German	18,444
en	English	252,289
es	Spanish	41,704
fr	French	16,992
hi	Hindi	27,132
id	Indonesian	177,867
ja	Japanese	1,274
ko	Korean	21,009
mr	Marathi	265
pt	Portuguese	21,540
ru	Russian	69,849
tr	Turkish	10,599
ur	Urdu	38,250
vi	Vietnamese	84,752
zh	Mandarin Chinese	25,159

| Total | | 821,361 |

Loading

from datasets import load_dataset, Audio

# Select a language; use "all" for the complete dataset.
ds = load_dataset("psdn-ai/CROWS-multilingual", "en",
                  split="full", streaming=True)

# Inspect metadata and retrieve the original WAV bytes without decoding.
ds = ds.cast_column("audio", Audio(decode=False))
row = next(iter(ds))
print(row["profile_id"], row["language"], row["prompt_id"])
assert row["audio"]["path"] == row["file_name"]

# Filter by any exposed metadata column.
selected = ds.filter(lambda row: row["prompt_id"] == "YOUR_SCRIPT_ID")

Use a dataset commit hash in revision= to reproduce an exact release. Streaming filters scan records; they do not imply an indexed random-access lookup. Audio decoding dependencies depend on the installed Hugging Face Datasets version.

Selection, validation, and splits

Every included recording corresponds to one unique ID in the supplied passing_files.csv; that inclusion list defines this release's selection. Metadata is parsed from source filenames of the form <profile_id>_<language>.<prompt_id>_<file_id>_<source_format>.

Packaging checks the S3 object version via its ETag, preserves its bytes, decodes every recording to check readability, validates archive member names and Parquet row counts, and records checksums. This validates packaging and readability, not the research labels or the selection criteria themselves.

The full split contains the complete release. This upload does not create experimental train/validation/test splits or add authenticity labels. Users should account for shared users, scripts, and related recordings when designing and reporting evaluation splits.

About Poseidon

Poseidon builds the training data for physical AI. Machines learn physical skill the way apprentices always have, by watching skilled people work, and that footage has to be created on purpose and to specification. Poseidon runs that supply chain: it recruits skilled people and real workplaces around the world, records their work to exact lab specifications, and refines the footage into model-ready datasets.

Research context and limitations

This release supports the publisher's multilingual audio research release. The metadata supplied for this upload does not include paper bibliographic information, detailed collection methodology, transcripts, authenticity labels, or numerical quality/spam scores. Consult the associated research documentation for those details; they are not inferred from filenames.

License, Access and Attribution

The dataset is licensed under Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0). See LICENSE and the official license.

When using this release, credit psdn-ai and link to CROWS Multilingual, identify the revision used, and indicate changes as required by the license.

Not what you're looking for? Request a custom dataset

Citation

If you use this dataset, please cite:

@misc{poseidonai2026crowsmultilingual,
  title        = {{CROWS Multilingual}},
  author       = {{Poseidon AI, Inc.}},
  year         = {2026},
  version      = {1.8},
  publisher    = {Hugging Face},
  url          = {https://huggingface.co/datasets/psdn-ai/CROWS-multilingual},
  howpublished = {\url{https://huggingface.co/datasets/psdn-ai/CROWS-multilingual}}
}

Profiles

The profiles folder provides 46,298 profiles with predicted synthetic submission proportions and distinct passing-language counts. Browse the profiles dataset configuration or download the three-column CSV, Parquet, or JSONL. Join audio metadata and profiles on profile_id. Reading pages are grouped by language count.