Waxal Datasets

The WAXAL dataset is a large-scale multilingual speech corpus for African languages, introduced in the paper WAXAL: A Large-Scale Multilingual African Language Speech Corpus.

Table of Contents

• Dataset Description
  • ASR Dataset
  • TTS Dataset
  • How to Use
• Dataset Structure
  • ASR Data Fields
  • TTS Data Fields
  • Data Splits
  • ASR v2 splits
• Dataset Curation
• Considerations for Using the Data
• Additional Information
• Citation

Dataset Description

The Waxal project provides datasets for both Automated Speech Recognition (ASR) and Text-to-Speech (TTS) for African languages. 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 datasets are collections acquired through partnerships with Makerere University, The University of Ghana, Digital Umuganda, Media Trust, Loud and Clear, and AIMS Senegal. Acquisition was funded by Google and the Gates Foundation under an agreement to make the dataset openly accessible. The Senegalese languages (Wolof and Pular) were provided by AIMS Senegal.

ASR Dataset

The Waxal ASR dataset is a collection of data in 19 African languages. It consists of approximately 1,250 hours of transcribed natural speech from a wide variety of voices. The 19 languages in this dataset represent over 100 million speakers across 40 Sub-Saharan African countries.

TTS Dataset

The Waxal TTS dataset is a collection of text-to-speech data in 17 African languages. It consists of over 180 hours of high-quality, single-speaker recordings reading phonetically balanced scripts.

How to Use

The datasets library allows you to load and pre-process your dataset in pure Python, at scale.

First, ensure you have the necessary dependencies installed to handle audio data. You will need ffmpeg installed on your system.

Google Colab / Ubuntu

sudo apt-get install ffmpeg
pip install datasets[audio]

macOS

brew install ffmpeg
pip install datasets[audio]

Windows Download and install from ffmpeg.org and ensure it's in your PATH.

pip install datasets[audio]

If you encounter RuntimeError: Could not load libtorchcodec, please ensure ffmpeg is correctly installed or check for compatibility between your torch, torchaudio, and torchcodec versions.

Loading ASR Data

To load ASR data for a specific language, specify the configuration name, e.g. sna_asr for Shona ASR data.

from datasets import load_dataset, Audio

# Load Shona (sna) ASR dataset
asr_data = load_dataset("google/WaxalNLP", "sna_asr")

# Access splits
train = asr_data['train']
val = asr_data['validation']
test = asr_data['test']

# Example: Accessing audio bytes and other fields
example = train[0]
print(f"Transcription: {example['transcription']}")
print(f"Sampling Rate: {example['audio']['sampling_rate']}")
# 'array' contains the decoded audio bytes as a numpy array
print(f"Audio Array Shape: {example['audio']['array'].shape}")

Loading TTS Data

To load TTS data for a specific language, specify the configuration name, e.g. swa_tts for Swahili TTS data.

from datasets import load_dataset

# Load Swahili (swa) TTS dataset
tts_data = load_dataset("google/WaxalNLP", "swa_tts")

# Access splits
train = tts_data['train']

Dataset Structure

ASR Data Fields

{
  'id': 'sna_0',
  'speaker_id': '...',
  'audio': {
    'array': [...],
    'sample_rate': 16_000
  },
  'transcription': '...',
  'language': 'sna',
  'gender': 'Female',
}

• id: Unique identifier.
• speaker_id: Unique identifier for the speaker.
• audio: Audio data.
• transcription: Transcription of the audio.
• language: ISO 639-2 language code.
• gender: Speaker gender ('Male', 'Female', or empty).

TTS Data Fields

{
  'id': 'swa_0',
  'speaker_id': '...',
  'audio': {
    'array': [...],
    'sample_rate': 16_000
  },
  'text': '...',
  'locale': 'swa',
  'gender': 'Female',
}

• id: Unique identifier.
• speaker_id: Unique identifier for the speaker.
• audio: Audio data.
• text: Text script.
• locale: ISO 639-2 language code.
• gender: Speaker gender.

Data Splits

For the ASR Dataset, the data with transcriptions is split as follows: * train: 80% of labeled data. * validation: 10% of labeled data. * test: 10% of labeled data.

The unlabeled split contains all samples that do not have a corresponding transcription.

The TTS Dataset follows a similar structure, with data split into train, validation, and test sets.

ASR v2 splits (speaker-disjoint)

The ASR half also ships a second set of splits intended for benchmarking, under data/ASR_v2/. They re-partition the same utterances so that no speaker appears in more than one split, meaning a model is always evaluated on voices it did not train on. The audio is untouched -- v2 is a re-labelling, not a new release of the recordings.

438,230 utterances, 2,242 hours. Gender is matched across splits for the 12 languages that carry gender labels. Per-language figures are in data/ASR_v2/metadata/split_comparison.csv, and the full provenance -- seed, source revision, every parameter -- in metadata/splits_manifest.json.

Already speaker-disjoint languages (Akan, Amharic, Oromo, Sidama, Tigrinya, Wolaytta)

Six languages were already speaker-disjoint in the original (v1) datasets and did not require a re-split:

• Amharic (amh), Oromo (orm), Sidama (sid), Tigrinya (tir), and Wolaytta (wal) (Digital Umuganda) already had 0.00% speaker leakage between train and validation/test.
• Akan (aka) (University of Ghana) already had 0.00% speaker leakage in its test split.

For these six languages, v2 directly points to the original dataset splits. You can load them directly via their v2 configs (aka_asr_v2, amh_asr_v2, orm_asr_v2, sid_asr_v2, tir_asr_v2, wal_asr_v2) or their original configs without the split map re-labelling procedure:

from datasets import load_dataset

# Direct loading for already speaker-disjoint languages (e.g. Amharic)
ds_test = load_dataset("google/WaxalNLP", "amh_asr_v2", split="test")
# Or equivalently using the original config:
# ds_test = load_dataset("google/WaxalNLP", "amh_asr", split="test")

Reading v2 (for re-partitioned languages)

Read the existing ASR configs and re-label each row from the split map:

import pandas as pd
from datasets import load_dataset

LANG, WANT = "sna", "test"
m = pd.read_csv(
    f"hf://datasets/google/WaxalNLP/data/ASR_v2/metadata/split_map/{LANG}.csv",
    dtype=str,
)
v2 = dict(zip(m["id"], m["v2_split"]))       # `id` is unique within a language

for v1_split in ("train", "validation", "test"):          # all three
    for row in load_dataset("google/WaxalNLP", f"{LANG}_asr",
                            split=v1_split, streaming=True):
        if v2.get(row["id"]) == WANT:
            ...

Three things worth knowing:

• Read all three original labelled splits. Utterances move between them: Acholi's v2 test set draws 397 of its 515 rows from the original train. Reading only the same-named split returns a fraction of the data, silently.
• Skip unlabeled. It carries no transcripts and takes no part in the re-split.
• Two utterances are unmapped -- kpo_149159 has no speaker_id and lin_9193 has an empty transcript. They are listed in metadata/excluded_rows.csv; v2.get() returns None for them.

Streaming bounds memory rather than download: filtering still transfers the language's labelled half, 1.6 GB (ach) to 12.7 GB (sid).

Dataset Curation

The data was gathered by multiple partners:

Considerations for Using the Data

Please check the license for the specific languages you are using, as they may differ between providers.

Affiliation: Google Research

Version and Maintenance

• Current Version: 2.0.0
• Last Updated: 03/2026

Citation

@article{waxal2026,
  title={WAXAL: A Large-Scale Multilingual African Language Speech Corpus},
  author={Anonymous},
  journal={arXiv preprint arXiv:2602.02734},
  year={2026}
}