Using a Large-Vocabulary, Continuous Speech Recognizer in a high-Volume application such as a commercial transcription service presents a different set of challenges and constraints than in a laboratory setting. We examine these differences with regard to acoustic model adaptation and find serious shortcomings in both the supervised and unsupervised approaches. We then examine a new method, semi-supervised adaptation, which overcomes the limitations of the other methods and can reduce recognition error rates by as much as 15% more than the reduction obtained through unsupervised adaptation.