CROSS-LINGUAL SPEECH RECOGNITION BETWEEN LANGUAGES FROM THE SAME LANGUAGE FAMILY

摘要

This paper presents a detailed analysis of how languages, members of different language subfamilies, influence word accuracy in the case of cross-lingual speech recognition. Cross-lingual speech recognition presents an efficient method for developing an automatic speech recognition system for ICT applications and services in the case of under-resourced languages, where the majority of Eastern and Central European languages can be classified. In the Romance language family experiments, Spanish was selected as the source and Romanian as the target language. The experiments were carried out using the Spanish SpeechDat(II) speech database and a dedicated Romanian speech database. The Spanish source automatic speech recognition system was trained using the MASPER system. A cross-lingual phoneme mapping approach based on expert knowledge was proposed for the Romanian-Spanish language pair. The Romanian cross-lingual speech recognition system achieved word accuracy of 80.48%.