Combination of a hidden tag model and a traditional n-gram model: a case study in czech speech recognition

A speech recognition system targeting high inflective languages is described that combines the traditional trigram language model and an HMM tagger, obtaining results superior to the trigram language model itself. An experiment in speech recognition of Czech has been performed with promising results. 1. Speech Recognition of Inflective Languages Inflective languages pose a hard problem in speech recognition due to two phenomena: highly inflective nature (causing data sparseness problem and excessive vocabulary growth), and free word order (causing the traditional speech recognition systems, such as n-gram Hidden Markov Models (HMMs) on word forms to be less accurate than for English). Specific methods targeting speech recognition of inflective languages have been already introduced in [1], [2] and [3]. The authors mainly focus on improving the language model by decomposing words from the vocabulary into stems and endings. This approach has mainly helped in reducing the size of the vocabulary of the speech recognizer reducing the WER slightly. 2. Combining Taggers with Language Models Tagger has been to our best knowledge first introduced as a speech recognition language model component in [4] without improving results over the baseline bigram model. The idea has been further explored in [5] where the author proposes the interpolation with a trigram model. P (W ) = λP (wi|wi−2, wi−1)+ (1 − λ)Q(wi|g(wi−2), g(wi−1)), (1) where g(wi) is the tagging function. The importance of formula (1) for languages with the data sparseness problem is that the new component Q can have enough evidence to give us reliable statistics about the word sequence W as the size of tag set tends to be much smaller then the size of the vocabulary itself. The problem with approach (1) is that the tagging function g(wi) depends on all words of the utterance (supposing that the tagging component is performed by an HMM tagger). The standard solution is to replace the probability Q by a new probability Q∗: Q∗(wi|w1, . . . , wi−1) = ∑ Q(wi|g1, g2)T (g(wi−2) = g2, g(wi−1) = g1) (2) The new probability T is the corresponding forward probability of the HMM tagger. The calculation of the forward probabilities is an seque words we do stand becom where proba are th that e inform us som the m the lis