查看文章

The unavailability of public datasets is the main hurdle for speech processing research targeting under-resourced languages. This paper reports the collection of a speech dataset comprising ten digits from the Kadazan language, which is one of the indigenous south-east Asian languages. Benchmark results for keyword spotting over the dataset using a convolutional neural network, have also been reported, with the benchmark model showing an average classification accuracy of 75.4% across multiple experiments using the dataset. Additionally, the dataset and implementation of the benchmark model have been made public, to facilitate replication and future research in the area of speech processing technologies for the Kadazan language.