The purpose of the work was to develop a speech signal processing method for the hearing pathologies correction based on psychoacoustically motivated transposition of high-frequency components of the signal spectrum to the low-frequency region with subsequent frequency-dependent amplification. To achieve this goal, several tasks related to the development of principles of frequency transposition in a speech signal were solved. The adjustment of the method is carried out according to the audiogram of a deaf person. For frequency transposition, source and target frequency bands are selected. The width of the source frequency band is fixed, while the width of the target band is adaptive. Spectrum transposition is performed only for consonants, the perception of which is more difficult for people with hearing loss. The classification of sounds (into vowel-consonant - pause classes) is implemented using one-layer neural network. The feature vector consists of: the zero crossing rate, short-term energy, short-term magnitude, normalized autocorrelation function and the first spectral moment. To preserve the naturalness of transposed sounds, the concept of equal loudness is used. To compensate for the attenuation in the perception of sound by a deaf person, a frequencydependent signal amplification based on an audiogram is used. The effectiveness of the proposed method was verified experimentally using hearing loss effect simulation. The experiment involved 10 people who were given to listen to the recordings passed through the hearing loss model, as well as recordings passed through the hearing loss model with subsequent correction (using proposed method). The results showed that using the proposed hearing correction method improves speech intelligibility on average by 6 %.
Stop gap removal using spectral parameters for stuttered speech signal
