ASR for low-resourced languages: Building a phonetically balanced Romanian speech corpus

The construction of automatic speech recognition (ASR) systems is fundamentally dependent on the speech corpus used to train the acoustic models. The speech corpus should be phonetically balanced to assure that the acoustic models are properly trained. This paper presents the design and development of the first phonetically balanced Romanian speech corpus. It describes all the language processing steps taken in order to obtain a proper set of phrases, discusses some important aspects regarding Romanian phonetics and emphasizes the phrase selection mechanism.