In this paper we describe a real-time speech recognition system developed for colloquial Iraqi Arabic. This system is currently used in our speech-to-speech translation system configured for bi-directional communication in English and Iraqi on a laptop. We present experimental results on Iraqi utterances from different speech-to-speech translation domains, and analyze the usefulness of acoustic and language modeling data from different domains. We highlight the improvements obtained by modeling techniques that are language-independent, such as lattice-based discriminative training and domain-biased language model interpolation. In addition, we report on initial experiments we have performed to address specific challenges posed by Iraqi for speech recognition such as absence of short vowels and multiple forms of glottal stop, or the hamza, in the written form.
Cite as: Saleem, S., Prasad, R., Natarajan, P. (2006) Colloquial Iraqi ASR for speech translation. Proc. Interspeech 2006, paper 1771-Wed1CaP.9, doi: 10.21437/Interspeech.2006-456
@inproceedings{saleem06_interspeech,
author={Shirin Saleem and Rohit Prasad and Prem Natarajan},
title={{Colloquial Iraqi ASR for speech translation}},
year=2006,
booktitle={Proc. Interspeech 2006},
pages={paper 1771-Wed1CaP.9},
doi={10.21437/Interspeech.2006-456}
}