Quality measurements for mobile data collection in the developing world

Abstract

The collection of speech data suitable for speech technology development is a challenge for under-resourced languages. Factors such as cost, availability of mother-tongue speakers and vast geographic distances call for techniques to optimise the data collection process in order to reduce re-collection of data. The use of mobile devices facilitate remote speech data collection. Although mobile (and remote) data collection addresses the challenging factors mentioned above, the environment is still less controlled than in the case of laboratory or studio-based recordings. In this paper we firstly revisit semi-realtime, basic quality control checks as implemented on available mobile-based speech data collection software (Woefzela). In addition, we introduce a quality control technique that uses speech duration estimation to validate the acoustic quality of the speech samples. We compare both techniques with manual verifications.