BENYO-S2ST-Corpus-1

Bilingual English-to-Yorùbá Speech-to-Speech Translation Corpus Version 1 (BENYO-S2ST-Corpus-1) is based on a hybrid architecture we developed for curating large scale dataset for direct speech-to-speech translation(S2ST) at a reduced cost. We leveraged non speech-to-speech Standard Yorùbá(SY) real-time recordings in the YORULECT Corpus and synthetic Standard English(SE) transcripts with their corresponding audios generated using pre-trained AI models (i.e. Facebook NLLB and MMS). 
We also developed an audio augmentation algorithm named AcoustAug based on three latent acoustic features (i.e. pitch, volume and speed) to generate augmented audios from the raw audios of the two languages. This approach for developing S2ST corpus is intended to address the limitations of existing datasets for English-to-Yorùbá language pairs. By employing hybrid data curation approach, the corpus aims to provide large scale, high-quality, and scalable digital resources for researchers and developers working on Yorùbá speech technologies and/or bidirectional English-Yorùbá speech pairs.

The BENYO-S2ST-Corpus-1 have 12,032 audio samples per language, which gives a total of 24,064 sample size. The minimum, maximum, average and total durations for English audios are 1.12s, 14.85s, 5.33s and 64,131.71s (i.e 17.81hours) respectively whereas the values for Yorùbá audios are 1.13s, 16.62s, 7.00s and 84,201.80(i.e. 23.39hours). Thus, the total audio duration for the two languages is 41.20 hours. This size is quite significant, given that existing high-to-low-resource S2S pairs have <20hours of parallel audios. For instance, the total duration for Yorùbá audios in FLEURS is approximately 15hours and it is less than 10hours in Common Voice (Conneau et al. 2023; Jia et al. 2022). Furthermore, the released audio pairs from the 2023 International Conference on Spoken Language Translation (IWSLT) S2ST Challenge for English-to-low resource target languages contain 17hours of English–Farsi,  8hours of English–Indonesian, 3hours of English–Catalan and 4hours of English–Vietnamese (Salesky et al., 2023).

Beyond building S2ST models, BENYO-S2ST-Corpus-1 can be used to build pre-trained models or improve existing ones for either of the languages(most especially Yorùbá) towards other downstream tasks such as TTS, S2T, ASR and NMT. The corpus architecture can further be leveraged by researchers/developers to curate corpora for multilingual-to-low resource African languages in order to bridge the huge digital divides in translations among high and low resource languages.

The files that are made publicly available for BENYO-S2ST-Corpus-1 are:


i) BENYO-S2ST-Corpus-1-Augmented-English-Audio-12k-Samples


ii) BENYO-S2ST-Corpus-1-Augmented-Yoruba-Audio-12k-Samples


iii) BENYO-S2ST-Corpus-1-Transcripts&Filenames-Metadata.xlsx

A paper on this corpus is available on ArXiv at - https://arxiv.org/abs/2507.09342

The work is funded through the 2024 Google Academic Research Award (GARA) for Society Centered Artificial Intelligence (SCAI) to Emmanuel Adetiba on the research project titled - A Direct Speech-to-Speech Model for English-to-Yoruba Translation Towards Bridging Language Barriers in Public Health Education Outreaches(https://bit.ly/3PQj7fq).

CONTACTS:

[email protected], [email protected]