# Multilingual Spoken Words Corpus (MSWC)

## Facts

- Provider: Mark Mazumder, Sharad Chitlangia, Colby Banbury et al. (MLCommons, Harvard)
- License: CC-BY-4.0 (https://huggingface.co/datasets/MLCommons/ml_spoken_words)
- Size: 1,333 files
- Stored at: Hugging Face, https://huggingface.co/datasets/MLCommons/ml_spoken_words
- Access: public, readable without an account
- Category: Speech & Audio
- Source: https://datasets-benchmarks-proceedings.neurips.cc/paper/2021/file/fe131d7f5a6b38b23cc967316c13dae2-Paper-round2.pdf
- Page: https://scigantic.com/archive/6dcf302f-db6e-539b-bd80-89e8cd63c705
- Keywords: Speech & Audio, Hugging Face

## About

The Multilingual Spoken Words Corpus is a large audio dataset of spoken words in 50 languages, built for keyword spotting and spoken term search. It contains more than 340,000 keywords and about 23.4 million one-second spoken examples, over 6,000 hours in total. The examples were generated by applying forced alignment to crowd-sourced sentence-level recordings from Mozilla Common Voice to obtain per-word timings and then extracting each word; all alignments are included. Use cases include on-device wake-word and keyword-spotting models, few-shot keyword spotting in low-resource languages and spoken term detection. This repository stores the audio in two formats, 16 kHz wav and 48 kHz opus, under data/wav and data/opus with one folder per language (50 each), and train/dev/test split files under data/splits (measured 2026-11-14; the Hub reports about 930 GB). Configurations are named '{lang}_{format}', for example datasets.load_dataset('MLCommons/ml_spoken_words', 'tt_wav'). Published by MLCommons under CC BY 4.0 for academic research and commercial applications; cite Mazumder et al., NeurIPS 2021 Datasets and Benchmarks.

## Load the data

```python
from huggingface_hub import hf_hub_download
path = hf_hub_download(repo_id="MLCommons/ml_spoken_words", filename="data/opus/ar/dev/audio/0.tar.gz", repo_type="dataset")
print(path)
```

## Data structure

Example file: `data/opus/ar/dev/audio/0.tar.gz`

## Open it

- In a hosted notebook with the data mounted: https://scigantic.com/archive/6dcf302f-db6e-539b-bd80-89e8cd63c705
- Full record as JSON (schema card included): https://api.scigantic.com/api/archives/6dcf302f-db6e-539b-bd80-89e8cd63c705
- Over MCP: `get_schema_card` with id `6dcf302f-db6e-539b-bd80-89e8cd63c705` at https://api.scigantic.com/mcp

## Related topics

- [Speech & Audio datasets](https://scigantic.com/topics/speech-and-audio)
