Teacher practice is normally assessed by observers who watch classes or videos of classes. Here, we analyse an alternative strategy that uses text transcripts and a support vector machine classifier. For each one of the 710 videos of mathematics classes from the 2005 Chilean National Teacher Assessment Programme, a single 4-minute slice was selected randomly and manually transcribed. The texts were then rated by two mathematicians and two teachers on the basis of the presence/absence of eight categories of contents and 12 categories of teacher practices. Then for each rater and category a classifier was built using a support vector machine algorithm trained with his/her ratings on a subset of texts. The agreements between the classifiers and the corresponding raters computed for texts belonging to an independent subset were better than the agreements between the raters. Next, we trained classifiers by using ratings created by raters who watched the videos rather than analysing the transcripts. Again, the agreements between classifiers and the raters were satisfactory. Given that class assessment can be an extremely slow, error prone and tedious process, the proposed automatic classifier scheme could be a promising alternative for monitoring classes and building online performance support system of teacher practices. (Contains 2 figures and 5 tables.)