Methods for creating statistical language models (SLMs) for spoken Completely Automated Turing Tests for Telling Computers and Humans Apart (CAPTCHAs) are disclosed. In these methods, candidate challenge items including one or more words are automatically selected from a document corpus. Selected ones of the challenge items are articulated by a machine text-to-speech (TTS) system as candidate articulations. Those articulations are ranked based on a human listener score indicating whether a candidate articulation originated from a machine. The SLM is then trained to recognize machine TTS articulations according to those rankings, by using a subset of the plurality of candidate challenge items identified as machine articulations as a seed set.