Evaluating the effect of using different transcription schemes in building a speech recognition system for Arabic

Abstract

It is well-known that the Arabic language poses non-trivial issues for Automatic Speech Recognition (ASR) systems. This paper is concerned with the problems posed by the complex morphology of the language and the absence of diacritics in the written form of the language. Several acoustic and language models are built using different transcription resources, namely a grapheme-based transcription which uses non-diacriticised text materials, phoneme-based transcriptions obtained from automatic diacritisation tools (SAMA or MADAMIRA), and a predefined dictionary. The paper presents a comprehensive assessment for the aforementioned transcription schemes by employing them in building a collection of Arabic ASR systems using the GALE (phase 3) Arabic broadcast news and broadcast conversational speech datasets LDC (2015), which include 260 h of recorded material. Contrary to our expectations, the experimental evidence confirms that the use of grapheme-based transcription is superior to the use of phoneme-based transcription. To investigate this further, several modifications are applied to the MADAMIRA analysis by applying a number of simple phonological rules. These improvements have a substantial effect on the systems’ performance, but it is still inferior to the use of a simple grapheme-based transcription. The research also examined the use of a manually diacriticised subset of the data in training the ASR system and compared it with the use of grapheme-based transcription and phoneme-based transcription obtained from MADAMIRA. The goal of this step is to validate MADAMIRA’s analysis. The results show that using the manually diacriticised text in generating the phonetic transcription can significantly decrease the WER compared to the use of MADAMIRA diacriticised text and also the isolated graphemes. The results obtained strongly indicate that providing the training model with less information about the data (only graphemes) is less damaging than providing it with inaccurate information.