This dataset forms the the Africa ASR domain adaptation benchmark. The goal of the dataset is to enable building of ASR models for African languages that can adapt to domians outside the training data.

The benchmark is made up of 2 languages, Wolof and Akan.

The training dataset for Akan is made up of general purpose data while the test dataset is financial data. The training dataset for Wolof is composed of general purpose data while the test dataset ins composed of data from the agricultural domain.

Dataset Details

Dataset Description

The Africa ASR domain adaptation benchmark is an ASR datasets designed to evaluate how well ASR models perform domain specific datasets. While many state-of-the-art ASR models perform well on general purpose datasets, their performance declines when evaluated on domain specific datasets especially for low resource languages.

Curated by: Makerere AI Lab
Funded by: Gates Foundation
Shared by: Makerere AI Lab
Languages(s) (NLP): Wolof, Akan
License: Apache 2.0

Dataset Sources

The data for the train splits of the dataset was obtained from general purpose datasets while the test splits were obtained from domain specific datasets.

Uses

The dataset can be used to evaluate the domain adaptation capability of ASR models.

Direct Use

The dataset should be used to train and evaluate ASR models for domain adaptation.

Out-of-Scope Use

The dataset should not be used to re-identify the people that contributed the audios.

Dataset Structure

A typical data point comprises the path to the audio file and its sentence.

{
  'File No': 'Wolof_domain_adaptation_benchmark.mp3',
  'audio': {
              'path': 'wolof_domain_adaptation_benchmark.mp3',
              'array': array([-0.00048828, -0.000183111, -0.00137329, ..., 0.00079346, 0.00091553, 0.00085449], dtype=float32),
              'sampling_rate': 16000
            }
  'transcript': 'dew df '
}

Data Fields

File No (string): Unique identifier for each audio file.
Audio (dict): A dictionary containing the path to the audio file, the decoded audio array, and the sampling rate. Note that when accessing the audio column: dataset[0]["audio"] the audio file is automatically decoded and resampled to dataset.features["audio"].sampling_rate. Decoding and resampling of a large number of audio files might take a significant amount of time. Thus it is important to first query the sample index before the "audio" column, i.e. dataset[0]["audio"] should always be preferred over dataset["audio"][0].
Transcription (string): The text corresponding to the audio.

Data Splits

The speech data has been subdivided into train and test splits.

Data Loading Recommended by HuggingFace

The following are data preprocessing steps advised by the HuggingFace. They are accompanied by an example code snippet that shows how to put them into practice.

from dataset import load_dataset

ds = load_dataset ("asr-africa/African-ASR-Domain-Adaptation-Evaluation")

Dataset Creation

Curation Rationale

The dataset was curated to train and evaluate ASR models for domian adaptation. Most ASR models perform well for general purpose data but when evaluated on domain specific data, their performance declines.

Dataset Card Authors

Makerere AI Lab

Dataset Card Contact

[email protected]

Downloads last month: 21

You need to agree to share your contact information to access this dataset

Dataset Card for Africa ASR Domain Adaptation Benchmark Dataset