Hugging Face's logo

Datasets:
google
/
WaxalNLP

Tasks:
Automatic Speech Recognition
Text-to-Speech
Modalities:
Audio
Text
Formats:
parquet
Languages:
Acoli
Akan
Amharic
Size:
1M - 10M
ArXiv:
Tags:
audio
automatic-speech-recognition
text-to-speech
Libraries:
Datasets
Dask
Polars
License:
Dataset card Data Studio Files
xet
 Community
20
WaxalNLP / README.md
perrynelson's picture
perrynelson
Update README.md
d145d01 verified
preview code
|
raw
history blame
8.37 kB
metadata 
license:
  - cc-by-sa-4.0
  - cc-by-nc-4.0
  - cc-by-4.0
annotation_creators:
  - human-annotated
  - crowdsourced
language_creators:
  - creator_1
tags:
  - audio
language:
  - ach
  - aka
  - dga
  - dag
  - ewe
  - ful
  - kpo
  - lin
  - lug
  - mlg
  - mas
  - nyn
  - sna
  - sog
multilinguality:
  - multilingual
pretty_name: Waxal Dataset
task_categories:
  - automatic-speech-recognition
  - text-to-speech
  - text-to-audio
source_datasets:
  - UGSpeechData
  - DigitalUmuganda/AfriVoice
  - original
dataset_info:
  features:
    - name: id
      dtype: string
    - name: speaker_id
      dtype: string
    - name: transcription
      dtype: string
    - name: language
      dtype: string
    - name: gender
      dtype: string
    - name: audio
      dtype: audio
  config_name: all
  splits:
    - name: train
    - name: validation
    - name: test
    - name: unlabeled

Waxal ASR Dataset

Table of Contents

Dataset Description

The Waxal dataset is a collection of automated speech recognition (ASR) data in 14 African languages. It consists of approximately 1,250 hours of transcribed natural speech from a wide variety of voices, suitable for ASR. The 14 languages in this dataset represent over 100 million speakers across 40 Sub-Saharan African countries. The goal of this dataset's creation and release is to facilitate research that improves the accuracy and fluency of speech and language technology for these underserved languages, and to serve as a repository for digital preservation.

The Waxal dataset is a collection acquired through partnerships with Makerere University, The University of Ghana and Digital Umuganda. Acquisition was funded by Google and the Gates Foundation under an agreement to make the dataset openly accessible.

Provider Languages License
Makerere University Acholi, Luganda, Masaaba, Nyankole, Soga CC-BY-4.0
University of Ghana Akan, Ewe, Dagbani, Dagaare, Ikposo CC-BY-NC-4.0
Digital Umuganda Fula, Lingala, Shona, Malagasy CC-BY-4.0

How to Use

The datasets library allows you to load and pre-process your dataset in pure Python, at scale. The dataset can be downloaded and prepared in one call to your local drive by using the load_dataset function.

The following language configurations may be used:

all, ach, aka, dga, dag, ewe, ful, kpo, lin, lug, mlg, mas, nyn, sna, sog

To download the config, specify the language code, (all for all languages, <language> code for a specific language).

Downloading specific language data (e.g. Shona):

from datasets import load_dataset

# Ensure you have the audio dependencies installed:
# pip install datasets[audio]

# Load Shona (sna) dataset
shona_data = load_dataset("google/WaxalNLP", "sna")

# Access splits
train = shona_data['train']
val = shona_data['validation']
test = shona_data['test']
unlabeled = shona_data['unlabeled']

# The 'audio' column is automatically decoded when accessed.
# It returns a dictionary containing 'path', 'array', and 'sampling_rate'.
example = train[0]
audio_data = example['audio']

print(f"Transcription: {example['transcription']}")
print(f"Audio array shape: {audio_data['array'].shape}")
print(f"Sampling rate: {audio_data['sampling_rate']}")

Downloading ALL data (large):

from datasets import load_dataset

all_data = load_dataset("google/WaxalNLP", "all")

Dataset Structure

The Waxal ASR dataset is a collection of recordings of speakers in 14 African languages with a human-written transcription of each recording. Each data point consists of a recording of at least 15 seconds, with a transcript. Each data point includes speaker-id, name of the language (one of 14 languages), speaker age, speaker gender, and speaker environment (Indoor, Outdoor, Other, In a car, Office or Studio).

Data Instances

  • Size of ASR dataset: 1.7T
  • Number of Instances: 1,937,031
  • Number of Fields: 6
  • Labeled Classes: N/A (Each label is a manually written transcription of the audio, no classes apply)
  • Number of Labels: 224,767
  • Percentage labeled instances: 11.62%
  • Algorithmic Labels: 0
  • Human Labels: 224,767
  • Hours: 10960

Language Codes

The data entries are grouped by ISO 639-2 language codes. This is so that the audio has a single universal name according to international standards removing ambiguity for languages that have multiple names.

ach, aka, dga, dag, ewe, ful, kpo, lin, lug, mlg, mas, nyn, sna, sog

The dataset includes 14 African languages:

ASR Language ISO 639-2 Audio Files Transcribed Hours Untranscribed Hours Total Hours
Acholi ach 114,308 32.32 659.49 691.81
Akan aka 195,285 101.93 941.37 1,043.30
Dagaare dga 191,404 104.66 949.80 1,054.47
Dagbani dag 188,808 98.54 962.28 1,060.82
Ewe ewe 203,391 99.77 976.58 1,076.35
Fulani ful 100,827 124.24 403.21 527.45
Ikposo kpo 191,984 103.81 941.22 1,045.03
Lingala lin 100,226 101.53 415.61 517.14
Luganda lug 98,475 45.96 631.44 677.40
Malagasy mlg 101,183 182.51 333.71 516.22
Masaaba mas 116,102 48.82 645.09 693.90
Nyankole nyn 131,743 50.87 754.51 805.38
Shona sna 102,969 99.23 474.94 574.16
Soga sog 120,172 50.34 736.45 786.79
Total all 1,956,877 1,244.52 9,825.71 11,070.23

Data Fields

The data is structured as follows:

{
  'id': 'sna_0',
  'speaker_id': '2Eud8lyLlsMcciYhmlkwVRtBwi82',
  'audio': {
    'array': [...],
    'sample_rate': 16_000
  },
  'transcription': '<transcription | "">',
  'language': 'sna',
  'gender': 'Female',
}

Field descriptions:

  • id: (string) unique identifier for the record.
  • speaker_id: (string) unique identifier for every speaker.
  • audio: (Audio) audio data for each example as a sound array.
  • transcription: (string) the transcription of the audio file if labeled, otherwise empty string.
  • language: (string) language code for the language in ISO 639-2 format.
  • gender: (string) represents the gender of the speaker if present, ('Male', 'Female' or empty if not present).

Data Splits

Each language configuration contains four data splits:

  • train: Contains 80% of samples with transcriptions.
  • validation: Contains 10% of samples with transcriptions.
  • test: Contains 10% of samples with transcriptions.
  • unlabeled: Contains all samples that do not have a transcription.

The all configuration will load data from all languages, while other configurations (e.g., sna) will load data only for the specified language.

Dataset Curation

The data is a curation of data that was gathered by multiple partners into one collective data collection with a standard interface to make it more universally accessible and useful for model training.

The original data sources and providers are listed below:

Provider Data Corpus License
University of Ghana UGSpeechData CC BY-NC-ND 4.0
Digital Umuganda AfriVoice CC-BY 4.0
Makerere University Yogera Dataset CC-BY 4.0

Considerations for Using the Data

When using this data corpus please keep in mind that data from different providers may license their data differently. Please check the license for the specific languages that you are using to make sure it is fit for your purposes.

Affiliation: Google Research

Version and Maintenance

  • Current Version: 1.0.0
  • Last Updated: 01/2026
  • Release Date: 01/2026