Unsupervised Method for Improving Arabic Speech Recognition Systems

31st Pacific Asia Conference on Language, Information and Computation (PACLIC 31), pages 161–168 Cebu City, Philippines, November 16-18, 2017 Copyright c⃝2017 Mohamed Labidi, Mohsen Maraoui and Mounir Zrigui Unsupervised Method for Improving Arabic Speech Recognition Systems Mohamed Labidi LaTICE laboratory Unit of Monastir 5000 Monastir, Tunisia labidi8mohamed@gmail.c om Mohsen Maraoui Computational Mathematics Laboratory, Tunisia 5000 Monastir, Tunisia maraoui.mohsen@gmail.c om Mounir Zrigui LaTICE laboratory Unit of Monastir 5000 Monastir, Tunisia mounir.zrigui@fsm.rnu.tn Abstract One of the big challenges connected to large vocabulary Arabic speech recognition is the limit of vocabulary, which causes high out- of-vocabulary words. Also, the Arabic language characteristics are another challenge. These challenges negatively affect the performance of the created systems. In this work we try to handle these challenges by proposing a new unsupervised graph-base method. Finally, we have obtained a 4.6% relative reduction in the word error rate. Comparing our suggested method with other methods in the literature, it has given better results. Moreover, it has presented a major step towards solving this problem. In addition, it can be easily adaptable to other languages. 1 Introduction and state of the art One of the big challenges in speech recognition is how to cover all possible words by a speech recognition system. The vocabulary of a conventional large-vocabulary continuous speech recognition system is finite, and this vocabulary limits the terms that appear in speech transcriptions. The words that do not occur in the vocabulary of the recognizer are called “out-of- vocabulary” words. This problem is a perennial challenge for speech recognition, where the out- of-vocabulary words are badly recognized. A larger vocabulary for the automatic speech recognition system is not the solution, since language is in constant growth and new words are steadily enriching the vocabulary. In (Ng and Zue, 2000), an analysis of news text demonstrated that the vocabulary size would continue to grow as the dataset got larger. In other words, it was not possible to create single large vocabulary that would eliminate the out-of- vocabulary problem. Consequently, it was not possible to create a language model that would cover all the words of any language. Furthermore, under certain conditions, adding more words could compromise the recognition performance of words already in the vocabulary. According to (Logan et al., 2005), up to 10% of all query words in a typical application that used a word-based recognizer with large vocabulary could be out-of-vocabulary words. Of course it was possible to update the vocabulary of the Automatic Speech Recognition (ASR) systems by adding new words to the language model. However, as noted by (Logan et al., 2005), it could be difficult to obtain enough training data to train the language model for new words. Additionally, for most application scenarios, it would not be feasible to re-recognize spoken content once the initial transcription was generated, due to the high computation cost of the ASR process and the huge sizes of daily spoken content collections. For these reasons, the out-of-vocabulary problem was a formidable one. For the Arabic language, this …

Content truncated to 500 words