Accessibility of speech information in videos is a huge challenge for the hearing impaired, making a visual representation such as text subtitling essential. Unavailability of a good Automatic Speech Recognition (ASR) engine, makes automatic generation of text subtitles for resource deficient languages such as Indian languages, extremely difficult. Techniques to build such an ASR using audio and corresponding transcription in the form of broadcast news or audio books have been proposed; however, these techniques require transcriptions corresponding to the audio in editable text format, which are unavailable for resource deficient languages. In this chapter, a novel technique of building a sound-glyph database for a resource deficient language has been described. The sound-glyph database can be used effectively to subtitle videos in the same language script. Considering large volumes of data that need to be processed, we propose a parallel processing method in a multiresolution setup, harnessing the multi-core capacity of present day computers.
Recovering capitalization and punctuation marks for automatic speech recognition: Case study for Portuguese broadcast news
