Современные методы цифровой обработки речевых сигналов
- № 2 (42) 2017
Страницы:
2
–
13
Язык: русский
Аннотация
В статье рассматриваются проблемы компьютерной обработки речевых сигналов, методы параметрического представления отдельных компонент речи. На этапе предварительной параметризации (акустического и фонетического анализа) предлагается использовать методы корреляционного и спектрального анализа. По результатам параметризации применяются процедуры распознавания с использованием возможностей нейронных сетей и марковских цепей.
Нутқ сигналларни компьютерда замонавий қайта ишлаш методлари кўрсатилган. Биринчи босқичда нутқ сигналларни акустик – фонетик параметрлари аниқланади. Бунинг учун корреляцион ва спектрал усуллардан фойдаланилади. Иккинчи босқичда эса нейрон тармоқлар ва марков занжирлар ёрдамида нутқ элементларини ва сўзларни таниш жараёнлари ифодаланган.
In the studies of analysis and recognition of separate and concurrent speech, scientifically based approaches to parameterization, solving spectral analysis problems, creating algorithms for identifying informative features and methods for their recognition, and the formation of new digital images of speech elements are needed. For the Uzbek speech, these tasks are particularly relevant, since up to the present time these issues have not received due attention. Today, the problem of mastering speech recognition methods and technologies arises, developed for the German language group languages — English, German and French. Known recognition methods [1] should be adapted to the problems of creating models for processing Uzbek speech.When creating programs for Uzbek speech recognition, the problem is solved from scratch, since the main time and resources will go to the process of building a reliable acoustic model, assessing the complexity of the language model, creating standards of words and grammars of the language.The words that make up the sentences form the basis of speech, so the first stage of research of the automatic processing of Uzbek speech is the acousticphonetic processing of isolated words.When creating systems for automatic analysis of speech signals, the main technique is to represent the signal in the form of piecewise stationary fragments and to process these fragments using correlation and spectral analysis methods. At this stage of rocessing, numerical methods are used that make it possible to recognize individual phonemes (sounds), syllables and whole words with high accuracy. In the subsequent stages, methods of the statistical theory of pattern recognition-neural networks, Markov chains and the method of dynamic programming, are used for comparison with standards.
Нутқ сигналларни компьютерда замонавий қайта ишлаш методлари кўрсатилган. Биринчи босқичда нутқ сигналларни акустик – фонетик параметрлари аниқланади. Бунинг учун корреляцион ва спектрал усуллардан фойдаланилади. Иккинчи босқичда эса нейрон тармоқлар ва марков занжирлар ёрдамида нутқ элементларини ва сўзларни таниш жараёнлари ифодаланган.
In the studies of analysis and recognition of separate and concurrent speech, scientifically based approaches to parameterization, solving spectral analysis problems, creating algorithms for identifying informative features and methods for their recognition, and the formation of new digital images of speech elements are needed. For the Uzbek speech, these tasks are particularly relevant, since up to the present time these issues have not received due attention. Today, the problem of mastering speech recognition methods and technologies arises, developed for the German language group languages — English, German and French. Known recognition methods [1] should be adapted to the problems of creating models for processing Uzbek speech.When creating programs for Uzbek speech recognition, the problem is solved from scratch, since the main time and resources will go to the process of building a reliable acoustic model, assessing the complexity of the language model, creating standards of words and grammars of the language.The words that make up the sentences form the basis of speech, so the first stage of research of the automatic processing of Uzbek speech is the acousticphonetic processing of isolated words.When creating systems for automatic analysis of speech signals, the main technique is to represent the signal in the form of piecewise stationary fragments and to process these fragments using correlation and spectral analysis methods. At this stage of rocessing, numerical methods are used that make it possible to recognize individual phonemes (sounds), syllables and whole words with high accuracy. In the subsequent stages, methods of the statistical theory of pattern recognition-neural networks, Markov chains and the method of dynamic programming, are used for comparison with standards.