Intelligent speech recognition data acquisition for Maltese

Abstract

Automatic Speech Recognition is a difficult task for under-resourced languages such as Maltese, as large quantities of data are required for its development. This dissertation seeks to provide a solution to this issue by crowdsourcing speech recordings and devising ways of validating this data efficiently. Common Voice was used as a crowdsourcing platform, facilitating the collection of 11+hours of speech data since its launch for Maltese. For validation, phonological analysis was performed on the text prompts using a grapheme-to-phoneme tool. The results of this were then compared to the number of syllables and segments detected in the speech using syllable nucleus detection and unsupervised automatic phoneme segmentation. Syllable distance between recordings and prompts was seen to be an effective metric for validation down to distances as small as a single syllable. Segment distance was effective when faced with differences of a few syllables or more.