Reduction in tlequency resolving capacity of the auditory system due to spread of masking of frequency cornponents by neighboring frequency components degrades speech perception in cases of sensorineural hearing impairment. We have carried out experimental evaluation of splitting speech into two signals lby using a bank of critical band filters, in order to reduce the effect of spectral masking in the cochlea. The dichotically presented signals are perceptually integrated in the auditory cortex. Listening tests were carried out with vowel-consonantvowel and consonant-vowel syllables for twelve English consonants on five normal hearing subjects with simulai ion of sensorineural impairment done by adding white masking noise to the speech signal at various S". Significant improvements in recognition score were obtained under adverse listening conditian. Improvement in the reception of speech feature of voicing, place, and manner was observed in information transmission analysis.

Persons with sensorineural hearing impairment face a particular problem in view of decrease in frequency resolving capacity of the auditory system due to spread of spectral masking along the cochlear partition. Filtering speech signal by a filter bank and adding signals from alternate bands and presenting it to the two ears is likely to reduce the effect of spectral masking and thus improving the speech intelligibility. The scheme was implemented in real time processing for use as a binaural hearing aid. Processing for each channel is done by linear phase FIR filter having a magnitude response with pass bands corresponding to alternate critical bands, the magnitude response for the two channels being complemented. For both the filters, the magnitude response was approximated with 128 coefficients using frequency sampling technique of linear phase FIR filter design. The implementation was done on two TI/TMS320C50 based DSP boards, each having 14-bit ADC and DAC. Twelve English consonants were used for carrying out listening tests in vowel-consonant-vowel (VCV) and consonant-vowel (CV) syllables presented to six hearing impaired subjects with bilateral sensorineural hearing loss. Information transmission analysis of confusion matrices for various features show maximum improvement for the place feature.

Sensorineural hearing impaired listeners face a particular problem in view of decrease in frequency resolving capacity of ear due to spread of spectral masking along the cochlear partition. Filtering speech signal by bank of critical band filters and adding signals from alternate bands for presenting to the two ears, is likely to reduce this effect, and thus may help in improving the speech intelligibility. We have implemented this processing scheme with eighteen critical bands for experimental evaluation. Listening tests were carried out using twelve vowelconsonant-vowel and consonant-vowel nonsense syllables presented in quiet, on ten subjects with mild-to-very severe sensorineural hearing loss. The stimulus response confusion matrices were analyzed for obtaining recognition scores and information transmission. The relative improvement in recognition scores is up to 25 percent. Information transmission analysis indicated that the overall improvement is contributed by improvements in transmission of place and manner features. The mean response time has also been found to significantly decrease.

No abstract

No abstract

No abstract

The social interaction of human beings is many times influenced by non-verbal communication, especially facial expressions. In societal life face of a human being is mostly observed by surrounding people to know the inner feelings. Thus, face forms a significant source of expressing human emotions, typically categorized into surprise, anger, fear, disgust, sad and happy. In the variety of behavioral science fields, emotion recognition has a significant role to play. The present paper describes a system in which preprocessing is performed by median filtering. Before extracting features, the watershed segmentation is applied to get the required characteristics of an image. In this paper, the Gamma based Feature Extraction (GFE), Local Binary Pattern (LBP) and Histogram of Oriented Gradients (HOG) technique have been used for feature extraction. The LBP algorithm is additionally tested with and without application of gamma correction using GFE. Two classifiers, namely kNN and SVM, have been employed, and their performance is compared. kNN and SVM, being supervised classifiers, can aid in better accuracy with proper training

No abstract