Morphologically motivated word classes for very large vocabulary speech recognition of Finnish and Estonian

We study class-based n-gram and neural network language models for very large vocabulary speech recognition of two morphologically rich languages: Finnish Estonian. Due to morphological processes such as derivation, inflection compounding, the need be trained with sizes several millions word types. Class-based modelling is in this case a powerful approach alleviate data sparsity reduce computational load. For vocabulary, bigram statistics may not an optimal way derive classes. thus utilizing output analyzer achieve efficient show that classes can learned by refining smaller equivalence using merging, splitting exchange procedures suitable constraints. This type classification improve results, particularly when model training large. also extend previous analyses rescoring hypotheses obtained from recognizer models. despite fixed carefully constructed word-based some cases result lower error rates than subword-based unlimited