---
dataset_info:
  features:
  - name: audio
    dtype:
      audio:
        sampling_rate: 16000
  - name: text
    dtype: string
  - name: lang
    dtype: string
  - name: accent
    dtype: string
  splits:
  - name: train
    num_bytes: 3365528653
    num_examples: 12368
  - name: valid
    num_bytes: 413508457
    num_examples: 1546
  - name: test
    num_bytes: 399289965
    num_examples: 1546
  download_size: 4177821855
  dataset_size: 4178327075
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
  - split: valid
    path: data/valid-*
  - split: test
    path: data/test-*
license: cc-by-4.0
tags:
- language-identification
- spoken-language-identification
- speech
- low-resource
- accent-robust
pretty_name: cld-multi-dataset
task_categories:
- audio-classification
- automatic-speech-recognition
language:
- en
- zh
- id
- ms
- hi
---

# CLD — Multilingual (5-language) Speech Dataset

Speech dataset for **Convex Low-resource Accent-Robust Language Detection (CLD)**,
covering 5 languages chosen for a deliberately challenging classification
boundary. This is the **multiclass** division of the CLD data.

[![paper](https://img.shields.io/badge/paper-ICML%202026-blue.svg)](https://arxiv.org/abs/2605.23235)
[![code](https://img.shields.io/badge/code-GitHub-181717.svg?logo=github)](https://github.com/pilancilab/CLD)
[![pypi](https://img.shields.io/badge/pip-jaxcld-3775A9.svg?logo=pypi&logoColor=white)](https://pypi.org/project/jaxcld/)

## Dataset description

We curate a dataset of multilingual voice transcriptions across high-resource
languages and their low-resource sub-dialects. As a primary source of
transcription data we use the **Common Voice (v23)** dataset (Ardila et al., 2020).
We supplement this with several additional dialect datasets for regional speech
variance:

- **Singaporean English** from the **National Speech Corpus (NSC)** — the first
  Singapore English corpus — provided through the Info-communications and Media
  Development Authority (IMDA) of Singapore. Singlish is selected because studies
  show it incurs particularly high error rates during voice transcription
  (Fong et al., 2002).
- The **Lahaja** dataset (Sanket et al., 2024), a benchmark comprising 12.5 hours
  of Hindi from 132 speakers across 83 Indian districts.

We normalize and augment all audio files via time stretching, volume gain, pitch
shift, and augmented background noise with **MUSAN** (Snyder et al., 2015).

## Schema

Each split is a [`datasets`](https://huggingface.co/docs/datasets) `Dataset` with
columns:

| column   | type                              | description                          |
|----------|-----------------------------------|--------------------------------------|
| `audio`  | `Audio(sampling_rate=16000)` mono | the speech clip, 16 kHz mono         |
| `text`   | `string`                          | reference transcription              |
| `lang`   | `string`                          | ISO-639-1 language code              |
| `accent` | `string`                          | accent / dialect label               |


## Multiclass setup

For the multiclass classification task we select 5 languages: **English, Chinese,
Indonesian, Malay, Hindi**. This selection establishes a challenging classification
boundary, since these languages share linguistic and geographical proximity — such
regional influences often cause misidentification (e.g. Singaporean English is
frequently confused with Malay or Indonesian). To maintain a low-resource setting
we curate ~16,000 training samples across these 5 languages, incorporating 24 unique
accents (~3,200 samples per language, ~666 per accent), with an 80-10-10
train/test/validation split.

- **Languages (5):** `en` (English), `zh` (Chinese), `id` (Indonesian), `ms` (Malay), `hi` (Hindi)
- **Splits:** `train` / `valid` / `test`

## How to use

```python
from datasets import load_dataset

ds = load_dataset("williamhtan/cld-multi-dataset")
print(ds)
sample = ds["test"][0]
print(sample["lang"], sample["text"])
audio = sample["audio"]          # {"array": np.ndarray, "sampling_rate": 16000}
```

## Citation

If you use this dataset, please cite the CLD paper (ICML 2026) and the underlying
corpora: Common Voice (Ardila et al., 2020), the National Speech Corpus (IMDA),
Lahaja (Sanket et al., 2024), and MUSAN (Snyder et al., 2015).