Sigita Laurinciukaite’s scientific contributions

This paper presents the development of Lithuanian continuous speech corpus LRN 1 (Lithuanian Radio News, version 1). The corpus was developed from speech corpus LRN 0.1 by increasing the duration of speech corpus (it lasts 20 hours 50 minutes). The major improvement of speech corpus LRN 1 was a development of time-aligned word level annotations of speech signals. Time-aligned word level annotations of speech signals were obtained after a two-Stage process: automatic realignment of acoustic models of phonemes and subsequent manual correction of annotations. The improvement of the corpus is useful for constructing and evaluating speaker-independent continuous speech recognition systems and for linguistic research.

This paper describes a framework for making up a set of syllables and phonemes that subsequently is used in the creation of acoustic models for continuous speech recognition of Lithua- nian. The target is to discover a set of syllables and phonemes that is of utmost importance in speech recognition. This framework includes operations with lexicon, and transcriptions of records. To fa- cilitate this work, additional programs have been developed that perform word syllabification, lexi- con adjustment, etc. Series of experiments were done in order to establish the framework and model syllable- and phoneme-based speech recognition. Dominance of a syllable in lexicon has improved speech recognition results and encouraged us to move away from a strict definition of syllable, i.e., a syllable becomes a simple sub-word unit derived from a syllable. Two sets of syllables and phonemes and two types of lexicons have been developed and tested. The best recognition accu- racy achieved 56.67% ±0.33. The speech recognition system is based on Hidden Markov Models (HMM). The continuous speech corpus LRN0 was used for the speech recognition experiments.

This paper presents a technique and experiments on choosing a set of mixed-duration modeling units representing language-specific phonetic details based on the analysis of available training data. A case study for Lithuanian speech recognition is presented. Lithuanian language defines the following major phonetic features: linguistic stress, consonant softness, vowel duration, and compound phones. Also, syllables or words could be chosen as modeling units. Incorporating all linguistic features into base phone set or using syllable models explodes the number of triphones used for accurate acoustic modeling. Therefore only those phones with phonetic features that have enough training samples are defined as separate phonetic units. The next problem is the choice of appropriate model complexity, which also depends on the available training data. Simple algorithms are proposed for forming dynamic base phone set and choosing complexity for separate models. Experiments on 10 hours of prepared speech from Lithuanian Radio news are conducted and indicate that the proposed modeling methodology allows significant improvement of recognition accuracy.

This paper presents a technique and experiments on choosing a set of mixed-duration modeling units representing language-specific phonetic details based on the analysis of available training data. A case study for Lithuanian speech recognition is presented. Lithuanian language defines the following major phonetic features: linguistic stress, consonant softness, vowel duration, and compound phones. Also, syllables or words could be chosen as modeling units. Incorporating all linguistic features into base phone set or using syllable models explodes the number of triphones used for accurate acoustic modeling. Therefore only those phones with phonetic features that have enough training samples are defined as separate phonetic units. The next problem is the choice of appropriate model complexity, which also depends on the available training data. Simple algorithms are proposed for forming dynamic base phone set and choosing complexity for separate models. Experiments on 10 hours of prepared speech from Lithuanian Radio news are conducted and indicate that the proposed modeling methodology allows significant improvement of recognition accuracy.