this is such a poor model, I've tried it before. Try yapsnap it has better word error rate and runs super fast on CPU. https://github.com/kouhxp/yapsnap/
Maybe there's a way to shrink it down given I only need to support a limited vocabulary domain...