ENHANCEMENT OF SPEECH INTELLIGIBILITY USING SPEECH TRANSIENTS EXTRACTED BY A WAVELET PACKET-BASED REAL-TIME ALGORITHM

摘要: Studies have shown that transient speech, which is associated with consonants, transitions between consonants and vowels, within some an important cue for identifying discriminating speech sounds. However, compared to the relatively steady-state vowel segments of has much lower energy thus easily masked by background noise. Emphasis can improve intelligibility in noise, but methods demonstrate this improvement either identified manually or proposed algorithms cannot be implemented run real-time.We developed algorithm automatically extract real-time. The involves use a function, we term transitivity characterize rate change wavelet coefficients packet transform representation signal. function large positive when signal changing rapidly small steady state. Two different definitions one based on short-time other Mel-frequency cepstral coefficients, were evaluated experimentally, MFCC-based produced better results. extracted used create modified combining it original speech.To facilitate comparison our processed using researcher emphasize transients, three indices. indices are extent modification/processing method emphasizes (1) particular region (2) relative to, (3) onsets offsets formants formant. These very useful because they quantify differences signals difficult show spectrograms, spectra time-domain waveforms.The extraction includes parameters varied influence speech. best values these selected psycho-acoustic testing. Measurements noise testing showed was more intelligible than especially at high levels (-20 -15 dB). incorporation identifies boosts unvoiced into does not result additional improvements.