Informatica

Conventional large vocabulary automatic speech recognition (ASR) systems require a mapping from words into sub-word units to generalize over the words that were absent in the training data and to enable the robust estimation of acoustic model parameters. This paper surveys the research done during the last 15 years on the topic of word to sub-word mappings for Lithuanian ASR systems. It also compares various phoneme and grapheme based mappings across a broad range of acoustic modelling techniques including monophone and triphone based Hidden Markov models (HMM), speaker adaptively trained HMMs, subspace gaussian mixture models (SGMM), feed-forward time delay neural network (TDNN), and state-of-the-art low frame rate bidirectional long short term memory (LFR BLSTM) recurrent deep neural network. Experimental comparisons are based on a 50-hour speech corpus. This paper shows that the best phone-based mapping significantly outperforms a grapheme-based mapping. It also shows that the lowest phone error rate of an ASR system is achieved by the phoneme-based lexicon that explicitly models syllable stress and represents diphthongs as single phonetic units.

The performance of an automatic speech recognition system heavily depends on the used feature set. Quality of speech recognition features is estimated by classification error, but then the recognition experiments must be performed, including both front-end and back-end implementations. We propose a method for features quality estimation that does not require recognition experiments and accelerate automatic speech recognition system development. The key component of our method is usage of metrics right after front-end features computation. The experimental results show that our method is suitable for recognition systems with back-end Euclidean space classifiers.

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 Lithuanian. 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 facilitate this work, additional programs have been developed that perform word syllabification, lexicon 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 accuracy 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.

There is presented a technique of transcribing Lithuanian text into phonemes for speech recognition. Text-phoneme transformation has been made by formal rules and the dictionary. Formal rules were designed to set the relationship between segments of the text and units of formalized speech sounds – phonemes, dictionary – to correct transcription and specify stress mark and position. Proposed the automatic transcription technique was tested by comparing its results with manually obtained ones. The experiment has shown that less than 6% of transcribed words have not matched.

The development of Lithuanian HMM/ANN speech recognition system, which combines artificial neural networks (ANNs) and hidden Markov models (HMMs), is described in this paper. A hybrid HMM/ANN architecture was applied in the system. In this architecture, a fully connected three‐layer neural network (a multi‐layer perceptron) is trained by conventional stochastic back‐propagation algorithm to estimate the probability of 115 context‐independent phonetic categories and during recognition it is used as a state output probability estimator. The hybrid HMM/ANN speech recognition system based on Mel Frequency Cepstral Coefficients (MFCC) was developed using CSLU Toolkit. The system was tested on the VDU isolated‐word Lithuanian speech corpus and evaluated on a speaker‐independent ∼750 distinct isolated‐word recognition task. The word recognition accuracy obtained was about 86.7%.