Some additional information: * No corpus of existing data * Client is open to leveraging mturk for training, at least initially * Fast latency is important,

Some additional information: * No corpus of existing data * Client is open to leveraging mturk for training, at least initially * Fast latency is important, but something on the order of ~5s is probably ok for v1 * ~85% "accuracy" is ok for v1