Unsupervised language model adaptation methods for spontaneous speech

In this paper we examine the performance of three different unsupervised language model adaptation schemes applied to speech recognition of spontaneous speech lecture presentations. Two of the schemes have been described previously in the literature while the third is a variation of one of the other two schemes. All three schemes are based on a combination of word -gram and class -gram models and use an initial transcription hypothesis to adapt the parameters of the class model. In each case the adapted class model is linearly interpolated with the baseline word -gram model and the combination is then applied in a subsequent recognition step. One of the schemes described also contains an element of domain adaptation in which the transcription hypothesis is also used to determine the interpolation weights of several class models each of which is built on automatically derived clusters of presentations. We also investigate multi-pass adaptation for each scheme and show this gives additional improvements in performance. Relative improvements in word error rate of up to 12.5% (3.4% absolute) are obtained on a held-out test set with the best adaptation scheme.