Bangla-Align: A forced-aligning toolkit for annotating Bangla speech

Abstract

This paper presents Bangla-Align, a toolkit for force-aligning Bangla speech. Forced-aligners are widely used tools for annotating speech at the phone level which greatly aids in increasing the processing power. While forced-aligning tools are freely available for a number of languages including, English, French, Spanish, etc., there is no aligner to process Bangla speech. Bangla-Align is built on top of the Montreal-forced-aligner (McAuliffe et al., 2017) with some Python scripts, which takes the audio recordings (.wav format) and their transcriptions (Text Grid format) as input and then returns Text Grids with phoneme-level annotations for phonetic/acoustic analyses of the audio data. The aligner uses a rule-based and continually developing phoneme dictionary. The aligner currently runs on Linux operating system only via command line interface. Bangla-Align will facilitate phonetic/acoustic research involving Bangla data.

Key words: Bangla-align, Bangla speech, forced aligner, align performance