Advances in Arabic broadcast news transcription at RWTH

Abstract

This paper describes the RWTH speech recognition system for Arabic. Several design aspects of the system, including cross-adaptation, multiple system design and combination, are analyzed. We summarize the semi-automatic lexicon generation for Arabic using a statistical approach to grapheme-tophoneme conversion and pronunciation statistics. Furthermore, a novel ASR-based audio segmentation algorithm is presented. Finally, we discuss practical approaches for parallelized acoustic training and memory efficient lattice rescoring. Systematic results are reported on recent GALE evaluation corpora.

A Novel Methodology to Extract Voice Signal Features

“…Many developments have been done in the field of voice recognition [9] and mostly all the developed proposed methods are based on: Voice analysis, feature extraction, modeling and matching [10]. These methods [11], [12], [13] and [14] are mostly suffering from the efforts needed to perform the process of voice recognition and the lowest recognition ratio. A method proposed in [15] is a good example of one of the current development which is based on extracting voice features based on voice analysis to calculate some parameters such as estimated population (mu), dynamic range, peak factor, Power spectral density, and zero crossing rate.…”

Phonetic pronunciations for arabic speech-to-text systems

“…The first two bnad06 and bcad06 were defined by BBN technology and consist of about 3 hours of BN and BC data respectively (collected during Dec05-Jan06). These test sets comprise com- 3 For the SPron system this only distinguished between a silence model being at the end of a word or not. …”

“…As part of the training process it is necessary to obtain pronunciations for words that can not be handled by Buckwalter [4,3]. A series of rules were automatically generated from a 250K Buckwalter derived phonetic dictionary.…”

“…Minimum Phone Error (MPE) discriminative training was used to train all the acoustic models. Pronunciation probabilities were used for the phonetic systems 3 . The N-gram language models were trained on data from 22 Arabic sources, including the acoustic transcriptions.…”

“…Recently there has been much interest in the problems associated with transcribing Arabic audio [1,2,3]. There are a number of issues to be addressed for success due to the nature of the Arabic language.…”