Voice centric interfaces have become widely available with the evolution of mobile phones. Advancements in recognition algorithms, increased computation power and networking capacities have enabled this growth. Applications like Siri are steadily gaining popularity with their hands free features. They can send data over to servers where sophisticated algorithms provide exceptionally accurate recognition results.
However in countries like India, where over 75 percent of the population has no internet, these applications are not suitable. Privacy is also a huge concern as these systems can share voice data with third parties. While open-source embedded systems like PocketSphinx are present, they require large amounts of training data which is usually not available for unpopular languages and is hard to collect and annotate.
In this work, we implement a portable word based embedded speech recognition system for mobile phones, along with a graphical user interface for training on desktop, and a demonstrative voice controlled assistant for android.
