Phoneme set selection for russian speech recognition

In this paper, we describe a method for phoneme set selection based on combination of phonological and statistical information and its application for Russian speech recognition. For Russian language, currently used phoneme sets are mostly rule-based or heuristically derived from the standard SAMPA or IPA phonetic alphabets. However, for some other languages, statistical methods have been found useful for phoneme set optimization. In Russian language, almost all phonemes come in pairs: consonants can be hard or soft and vowels stressed or unstressed. First, we start with a big phoneme set and then gradually reduce it by merging phoneme pairs. Decision, which pair to merge, is based on phonetic pronunciation rules and statistics obtained from confusion matrix of phoneme recognition experiments. Applying this approach to the IPA Russian phonetic set, we first reduced it to 47 phonemes, which were used as initial set in the subsequent speech model training. Based on the phoneme confusion results, we derived several other phoneme sets with different number of phonemes down to 27. Speech recognition experiments using these sets showed that the reduced phoneme sets are better than the initial phoneme set for phoneme recognition and as good for word level speech recognition.