SOREVA

SOREVA (Small Out-of-domain Resource for Various African languages) is a multilingual speech dataset designed for the evaluation of text-to-speech (TTS) and speech representation models in low-resource African languages. Comming from Goethe Institut intiative of collecting 150 samples(Audio and transcription) for about 49 africain languages and dialectes This dataset specifically targets out-of-domain generalization, addressing the lack of evaluation sets for languages typically trained on narrow-domain corpora such as religious texts.

SOREVA includes languages from across Sub-Saharan Africa, including:

• Standard languages:
  Afrikaans, Hausa, Yoruba, Igbo, Lingala, Kiswahili, isiXhosa, isiZulu, Wolof

• Dialectal & minor languages:
  Bafia, Bafut, Baka, Bakoko, Bamun, Basaa, Duala, Ejagham, Eton, Ewondo, Fe,
  Fulfulde, Gbaya, Ghamála, Isu, Kera, Kom, Kwasio, Lamso, Maka, Malagasy, Medumba,
  Mka, Mundang, Nda, Ngiemboon, Ngombala, Nomaande, Nugunu, Pidgin, Pulaar,
  Sepedi, Tuki, Tunen, Twi, Vute, Yambeta, Yangben, Yemba, Éwé

How to use & Supported Tasks

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.

For example, to download the Hindi config, simply specify the corresponding language config name (i.e., "ha_ng" for Nigeria):

from datasets import load_dataset
#Load a specific language (e.g., 'ha_ng' for Hausa Nigeria)
dataset = load_dataset("OlameMend/soreva", "ha_ng", split="test")

Load all languages together

from datasets import load_dataset 
dataset = load_dataset("OlameMend/soreva", "all", split="test")

1. Out-of-domain TTS & ASR model Evaluation

Getting Audio and Transcription

You can easily access and listen to audio samples along with their transcriptions:

from datasets import load_dataset
from IPython.display import Audio

#Load the dataset for a specific language, e.g., "ha"
soreva = load_dataset("OlameMend/soreva", "ha_ng", split='test' , trust_remote_code=True)

#Access the first example's audio array and sampling rate
audio_array = soreva[0]['audio']['array']        # audio data as numpy array
sr = soreva[0]['audio']['sampling_rate']         # sampling rate

#Print the corresponding transcription or use it for TTS inference for evaluation
print(soreva[0]['transcription'])

#Play the audio in a Jupyter notebook
Audio(audio_array, rate=sr)

Dataset Structure

We show detailed information the example configurations ewo_cm of the dataset. All other configurations have the same structure.

Data Instances

ewo_cm

• Size of downloaded dataset files: 14 MB

An example of a data instance of the config ewo_cm looks as follows:

{'path': '/home/mendo/.cache/huggingface/datasets/downloads/extracted/3f773a931d09d3c4f9e9a8643e93d191a30d36df95ae32eedbafb6a634135f98/cm_ewo_001.wav',
 'audio': {'path': 'cm_ewo/cm_ewo_001.wav',
  'array': array([-0.00518799, -0.00698853, -0.00814819, ..., -0.02404785,
         -0.02084351, -0.02062988]),
  'sampling_rate': 16000},
 'transcription': 'mbembe kidi',
 'raw_transcription': 'mbəmbə kídí',
 'gender': 0,
 'lang_id': 15,
 'language': 'Ewondo'}

Data Fields

The data fields are the same among all splits.

• path (str): Path to the audio file.
• audio (dict): Audio object including:
  • array (np.array): Loaded audio waveform as float values.
  • sampling_rate (int): Sampling rate of the audio.
  • path (str): Relative path inside the archive or dataset.
• transcription (str): Normalized transcription of the audio file.
• raw_transcription (str): Original non-normalized transcription of the audio file.
• gender (int): Class ID of gender (0 = MALE, 1 = FEMALE, 2 = OTHER).
• lang_id (int): Class ID of the language.
• language (str): Full language name corresponding to the lang_id.

Data Splits

Currently, as this is the first initiative, we only provide a test split containing approximately 150 audio samples.

Other splits such as train and validation are not included at this stage but are expected to be added through community contributions and continuous dataset development.

This initial test split allows evaluation and benchmarking, while the dataset will evolve to include more comprehensive splits in the future.

Dataset Creation

The data were collected by the Goethe-Institut and consist of 150 audio samples with corresponding transcriptions across 48 African languages and dialects.

Considerations for Using the Data

Social Impact of Dataset

This dataset is meant to encourage the development of speech technology in a lot more languages of the world. One of the goal is to give equal access to technologies like speech recognition or speech translation to everyone, meaning better dubbing or better access to content from the internet (like podcasts, streaming or videos).

Discussion of Biases

For all languages, only male voice is represented

Other Known Limitations

Certains transcript only contain single word instead of complete sentence; others line of transcription has deux sentences(variance) for the same audio

Additional Information

All datasets are licensed under the Creative Commons license (CC-BY).

Citation Information

Contributions

Thanks to @LeoMendo for adding this dataset.