Image approach to voice recognition

Systems for user verification are constantly developed, for which we need novel methods and approaches. In this article we present our research on the model for sound processing. Input signal is transformed by the use of Discrete Fourier Transform into spectrogram. In this way we receive an image visualization of the sound amplitude spectrum over time interval. For this we use a combination of heuristic approach and Convolutional Neural Network to search for significant features and recognize them. Presented results of numerical experiments show high potential of this idea and possibility for further development.

Перевод пока недоступен