NCHLT Auxiliary speech data for ASR technology development in South Africa

Area of Science:

• Computational Linguistics
• Speech Technology
• Corpus Linguistics

Background:

• South Africa has 11 official languages, all currently under-resourced for Human Language Technology (HLT) development.
• The National Centre for Human Language Technology (NCHLT) project aimed to create essential speech and text resources for these languages.
• Previous NCHLT releases, like the 2014 Speech Corpus, did not encompass all collected data.

Purpose of the Study:

• To describe and release additional speech and text data collected during the NCHLT project.
• To provide valuable resources for advancing speech technology for under-resourced South African languages.
• To supplement existing corpora and further enable HLT development.

Main Methods:

• Speech data collection via a smartphone application during the NCHLT project.
• Compilation and release of auxiliary speech corpora, including transcriptions for each utterance.
• Data curation and organization for accessibility and usability in HLT research.

Main Results:

• Release of auxiliary speech corpora containing 20 to 170 hours of data per language.
• Inclusion of transcriptions for all utterances within the auxiliary datasets.
• Significant expansion of available speech resources for South Africa's official languages.

Conclusions:

• The newly released auxiliary corpora significantly contribute to addressing the resource scarcity for HLT in South Africa.
• This data facilitates the development of speech technology for under-resourced languages.
• Further development of Human Language Technology for all 11 official South African languages is now more feasible.