Resources creation of Bengali for SPPAS

The development of HLT (Human Language Technology) tools inevitably involves the need for language resources. However, only a handful of languages possess such resources for free. This paper presents the development of such speech tools for the Bengali language. It focuses particularly on developing language resources for text normalisation, phonetization, forced-alignment and syllabification of Bengali. This aids in developing an automatic speech annotation system for predicting the pronunciation and alignment at phoneme, syllable and word levels of the speech with their corresponding orthographic text in this less-resourced language. Subsequently application of these forced-alignment tools on both read and spontaneous speech styles of Bengali is compared in order to evaluate the performance of the SPPAS (SPeech Phonetization Alignment and Syllabification) software. Spontaneous speech is filled with informal activities like laughter, noises and hesitations that make identification and alignment of phonetic entities with their corresponding transcription more problematic. The UBPA (Unit Boundary Positioning Accuracy) measure of the proposed forced-alignment system is 93.95% accurate in the case of read speech and 76.53% for spontaneous speech within a delta range of 40 ms. These newly created resources of Bengali have been integrated into the SPPAS software and distributed freely under the terms of public licenses.