Name recognition and synthesis can dramatically improve customer experience, disambiguate confusion, and make interactions more clear and personalized. Because they do not follow the same pronunciation rules as words, names are notoriously difficult to implement well. This presentation shows you how to avoid common pitfalls in name recognition and synthesis, discusses the problems associated with obtaining good data for name pronunciation, and describes best practices for developing good software models.

This talk discusses the behavior of two distinct user groups interacting with automated speech systems for collecting addresses. One group of callers was hyper-careful, speaking very slowly, with long pauses between words, often spelling out words and giving full ZIP+4 numbers. The second caller group was much less careful, spoke more quickly, and there was often side speech and background noise during the call. This presentation compares these two groups and describes how Novauris adjusted the design to accommodate their behavior.