The paper presents methods for evaluating the accuracy of alignments between transcriptions and audio recordings. The methods have been applied to the Spoken British National Corpus, which is an extensive and varied corpus of natural unscripted speech. Early results show good agreement with human ratings of alignment accuracy. The methods also provide an indication of the location of likely alignment problems; this should allow efficient manual examination of large corpora. Automatic checking of such alignments is crucial when analysing any very large corpus, since even the best current speech alignment systems will occasionally make serious errors. The methods described here use a hybrid approach based on statistics of the speech signal itself, statistics of the labels being evaluated, and statistics linking the two.