here is page 271 , down below

7.1 Analyze each of the errors in the incorrectly recognized transcription of "um the phone is I left the..." on page 271. For each one, give your best guess as to whether you think it is caused by a problem in signal processing, pronunciation modeling, lexicon size, language model, or pruning in the decoding search. Here is an example of alignments between a reference and a hypothesized utterance from the CALLHOME corpus, showing the counts used to compute the word error rate: This utterance has six substitutions, three insertions, and one deletion: WordErrorRate=100186+3+1=56% As of the time of this writing, state-of-the-art speech recognition systems were achieving around 20% word error rate on naturalspeech tasks like the National Institute of Standards and Technology (NIST)'s Hub4 test set from the Broadcast News corpus (Chen et al., 1999), and around 40% word error rate on NIST's Hub5 test set from the combined Switchboard, Switchboard-II, and CALLHOME corpora (Hain et al., 1999)