Building an Arabic news transcription system with web-crawled resources

This paper describes our efforts to build an Arabic ASR system with web-crawled resources. We first describe the processing done to handle Arabic text in general and more particularly to cope with the high number of different phonetic transcriptions associated to a typical Arabic word. Then, we present our experiments to build acoustic models using only audio data found in the web, in particular on the Euronews portal. To transcribe the downloaded audio we compare two approaches: the first uses a baseline trained on manually transcribed Arabic corpora, while the second uses a universal ASR system trained on automatically transcribed speech data of 8 languages (not including Arabic). We demonstrate that with this approach we are able to obtain recognition performances comparable to the ones obtained with a fully supervised Arabic baseline. keywords: Arabic, ASR, Morphological segmentation, Lightly supervised training, Under-resourced languages