Development of Spectro temporal Features of speech

Abstract

Speech and language development typically involves the ability to perceive the phonological structure of a language, relate this to the production of these sounds and finally produce the appropriate speech sounds for social communication. A deficiency in any of these processes could lead to language impairments notable in several neurodevelopment disorders like autism. Therefore, understanding acoustical basis of speech development can provide profound insights into the cognitive development of children. newline From a signal processing point of view, speech can be described as a time-pressure waveform. A wealth of research has shown that the auditory system acts as a frequency analyzer and the role of the frequency spectrum in encoding perceptual attributes of speech has been well established. Literature suggests that for a complete understanding of how various linguistic cues are encoded in a speech signal, it is important to adopt a spectro-temporal approach. Thus, spectro-temporal features of speech, being the phonological basis of perception in the auditory cortex, provide a relevant framework to study speech signal. newline In this thesis, a procedure to extract spectro-temporal features of speech has been provided. It is based on applying various filters on a spectrographic image to extract these features and determine their timescales. The statistics of these features of speech are also determined based on their timescales. Basically, two timescales were considered for this study; namely, long timescales of the order 100-500ms and short timescales of the order of 25-50ms. The long timescales represent supra-segmental features of speech like syllabicity, stress, tempo and rhythms. The short timescales represent segmental features like onset-rime units, phonemes, formants and their transitions. It is believed that these timescales represent different linguistic and phonological units. The proposed framework is then applied to a speech corpus consisting of 170 speech samples, where it is used to study development of speech production in normally developing children. The results show the presence of a significantly less number of features encoded at short timescales in children as compared to adults, but no significant difference is observed in long timescale features. The thesis also demonstrates the utility of this framework for studying speech impairments in children and adults with mild to moderate intellectual disabilities. It is evident from the results that many spectro-temporal features encoded at both the timescales are absent in subjects with intellectual disabilities; more prominently of shorter timescales. Finally, the differentiating power of spectro-temporal features is examined by training various classifiers using these features, which have yielded 95% success in classifying Intellectually Disabled. These results provide an initial step towards developing speech based early diagnostic tools and therapies newline newline