Last time I looked into OCR stuff I came to a similar conclusion (though I didn't implement anything back thne). It would be really nice to have "open source" models that had similar performance, without having to deal with the iphone cluster hackery.