Publications Repository - Gdańsk University of Technology

Much attention is given by researchers to the speech processing task in automatic speech recognition (ASR) over the past decades. The study addresses the issue related to the investigation of the appropriateness of a two-dimensional representation of speech feature spaces for speech recognition tasks based on deep learning techniques. The approach combines Convolutional Neural Networks (CNNs) and timefrequency signal representation converted to the investigative feature spaces. In particular, fractal dimension features of the signal were chosen for the time domain, and two feature spaces were investigated for the frequency domain, namely: frequency tracks obtained from the frequencies and amplitudes of the detected spectral peaks and the modified chromagrams. Both are constructed from a series of short-time Fourier transforms, which were computed along the window speech signal in the time domain. Due to the fact that deep learning requires a sufficiently large training set as the size of the corpus may significantly influence the outcome, thus for the data augmentation purpose, the created dataset was extended by adding various noise levels and mixed with the speech signal. In order to evaluate the applicability of implemented feature spaces for isolated word recognition task, three experiments were conducted: a 10-word, a 70-word, and a 111-word cases were analyzed.

Authors

• dr Grazina Korvel link open in new tab ,
• Gintautas Tamulevicus,
• Povilas Treigys,
• Jolita Bernataviciene,
• prof. dr hab. inż. Bożena Kostek link open in new tab