PRONUNCIATION VARIATION MODELING FOR ARABIC ASRs: A DIRECT DATA-DRIVEN APPROACH

Pronunciation variation is a major obstacle in improving the performance of Arabic automatic speech recognition (ASR) systems. This phenomenon alters the pronunciation spelling of words beyond their listed forms in the phonetic dictionary, leading to a number of out of vocabulary (OOV) word forms. This paper presents a direct data-driven approach to model pronunciation variations, in which the pronunciation variants are distilled from the training speech. The proposed method is based on adding the pronunciation variants to the ASR pronunciation dictionary as well as to the language model. We started with a baseline Arabic speech recognition system using Carnegie Mellon University (CMU) Sphinx3 speech recognition engine. The baseline is based on a 5.4 hour speech corpus of Modern Standard Arabic (MSA) broadcast news, with a phonetic dictionary of 14,234 canonical pronunciations. The baseline system achieves a WER of 13.39%. The proposed method identifies the variations in the phonetic transcription of the spoken words. The phonemic variants of words are then filtered and added with distinctive letter spellings in an expanded phonetic dictionary. In our experiment, 554 variants were added to the basic phonetic dictionary as new words. The artificially added words are used together with their sentences in the language model as well. Our results show that while the expanded dictionary alone did not add appreciable improvements, the WER is reduced by 2.04% when the variants are considered within the language model.