System for Speech Corpus Development

Many modern speech recognition systems now use neural networks. Although neural networks can produce acceptable results in some situations, there are a number of problems they cannot solve. Such speech recognition systems cannot properly recognize an unfamiliar dialect or speech with an accent or impairment. In addition, neural networks are a "black box", so that we cannot assess the correctness of their work or confidently know what we will get at the output, and can only observe the result. In this paper the acoustic-phonetic approach is discussed. The problem of speech segmentation and phoneme dataset collection with the description of speaker features is considered. The structure of a graph database for storing speech records, phonemes, description of speakers characteristics such as native language, dialect, presence of accent and speech disorders is presented. The speech corpus development system architecture is described.