Speech recognition datasets for low-resource Congolese languages

Area of Science:

• Computational Linguistics
• Speech Processing
• Low-Resource Language Technologies

Background:

• Large pre-trained Automatic Speech Recognition (ASR) models benefit from transfer learning, but require substantial data.
• Many low-resource languages lack sufficient data to fully leverage transfer learning.
• Benchmark corpora are essential for advancing ASR methods in data-scarce linguistic contexts.

Purpose of the Study:

• To introduce two novel benchmark corpora for low-resource languages in the Democratic Republic of the Congo.
• To facilitate the development of monolingual and multilingual ASR systems for Congolese languages.
• To enable inaugural benchmarking of speech recognition systems for Lingala and four other Congolese languages.

Main Methods:

• Creation of the Lingala Read Speech Corpus (4h labeled audio) with diverse speakers and accents.
• Compilation of the Congolese Speech Radio Corpus (741h unlabeled audio) from broadcast archives.
• Application of supervised learning and self-supervised learning techniques for model development and benchmarking.

Main Results:

• The developed corpora provide valuable resources for ASR research in low-resource settings.
• Successful inaugural benchmarking of speech recognition systems for Lingala.
• Development of the first multilingual ASR model for four Congolese languages, serving 95 million people.

Conclusions:

• The released datasets are crucial for advancing ASR research and development for underserved languages.
• These resources pave the way for improved speech recognition technologies in the Democratic Republic of the Congo.
• The study highlights the potential of transfer learning and novel corpora for low-resource ASR.