Actually, today's speech recognizers are really good and ones like Google's have accuracy >90%. So writing a speech recognizer ourselves with 50% accuracy and training it endlessly without ever reaching Google SR's accuracy sounded like a terrible 'reinvent-the-wheel' idea.

Instead, we wrote algorithms that further boosted the accuracy of the existing speech recognizer with machine-generated phonetic dictionaries. (You can think of it as standing on the shoulders of giants). The plan is to work with a couple of partners to train our machine learning system for their products to really bump up SR accuracy.

If you're building an app that's going to go live, send us a message, we'd be happy to work with you (for free!)