Stop Classification Using Desa - 1 High - Resolution Formant

Recent work has veriied that the second-formant frequency (F2) and its change in the vowel immediately preceding a stop consonant are usually suucient to discriminate between labial, palatal, and alveolar stops, even in the absence of the stop burst information. Informal listening tests using truncated samples indicate that humans can discriminate the three stops on the basis of the preceding vowel alone. Typical quasi-stationary analyses like LPC and DFT l-terbanks may not have suucient time-frequency resolution to detect the rapid F2 variations, and therefore a valuable source of stop classiication information is being overlooked. This paper shows the results of using the DESA-1 quadratic frequency estimator to determine the frequency and rate of change of F2. It is shown for diierent vowel environments that the DESA-1 algorithm can extract suucient information to classify stops from vocalic data. The performance is demonstrated to be superior to a formant tracker using a more conventional pitch-synchronous LPC analysis. The Discrete Energy Separation Algorithm (DESA-1) recently presented by Maragos, Kaiser, and Quatieri 1, 2] is based on the work of H. Teager. The DESA-1 algorithm provides a simple and elegant method of estimating the amplitude and frequency of a sinusoid subject to amplitude or frequency modulation. Using the notation of 2], the Teager energy-tracking operator is deened as c x(t)] 4 = _ x 2 (t) ? x(t) x(t) (1) where _ x(t) and x(t) are the rst and second time derivatives of x(t). Given a sinusoid with amplitude modulation a(t) and instantaneous frequency !(t), x(t) = a(t) cos(Z t 0 !()dd) (2) the Teager energy operator estimates the squared product of the instantaneous amplitude and frequency, under certain reasonable conditions (essentially, modulation frequencies must be slow with respect to the carrier). c x(t)] a 2 (t)! 2 (t) (3) Applying the operator to the derivative of the AM-FM signal yields c _ x(t)] a 2 (t)! 4 (t): (4) Clearly, the instantaneous amplitude and frequency may be found from the Teager energy estimates of the signal and its derivative by solving Equations (3) and (4) for a and !. In the discrete case, time derivatives may be approximated by time diierences. The discrete-time counterpart of the Teager operator c x(t)] becomes: d x(n)] 4 = x 2 (n) ? x(n ? 1)x(n + 1) (5) A discrete-time AM-FM sinusoidal signal having amplitude modulation a(n), carrier frequency c , modulation frequency m , and …