Hilbert-Huang Transform And Its Application In Feature Extraction Of Speech

Speech signal processing technology is one of the important means for the computer's intelligent interface and human-computer interaction. With the development of electronic computers and artificial intelligence, more and more speech coding, synthesis and recognition technology have been applied in people's life. And accurate extraction of the speech signal's feature parameters is the key of these speech signals processing technology.The speech signal is a complex non-linear, non-stationary signal, it makes the performance of traditional speech signal processing technology which based on linear and stationary systems theory difficult to have further improvement. This paper studies a signal processing method named Hilbert-Huang Transform which can effectively response the signal's non-linearity and non-stationary, the endpoint and stop criterion of the Hilbert-Huang Transform, estimations of instantaneous frequency and amplitude were analyzed, and presents some corresponding improvements. Pitch detection algorithm and formant extraction algorithm was studied used by the improved Hilbert-Huang Transform.At first, the endpoint and stop criterion of the Hilbert-Huang Transform, estimations of instantaneous frequency and amplitude have been studied in this paper, and three improvements have been proposed. The first one is combined with the characteristics of the speech signal, an segmentation and Endpoint Extension method based on speech signal is proposed; the second one is using the orthogonal property of IMF components, an stop criterion is proposed; by study the limitations of Hilbert demodulation method, put forward using the energy separation algorithm to estimate instantaneous frequency and instantaneous amplitude. Simulations results show that these methods have a good effect on inhibit the endpoint effects and modal mixing of Hilbert-Huang Transform.This paper has researched the pitch detection algorithm by using the improved Hilbert-Huang Transform. On the basis of Hilbert-Huang Transform's instantaneous energy, frequency weighted instantaneous energy is proposed. Compare with the Hilbert-Huang Transform's instantaneous energy, signal's frequency weighted instantaneous energy can not only reflect the instantaneous signal energy but also reflect signal's energy density. By using the speech signal's frequency weighted instantaneous energy glottal pulse can precisely locate in time, so precise pitch detection become possible. The experiment results show that comparing with the traditional instantaneous energy, this algorithm is more able to reflect the signal's real instantaneous energy information, so that they can improve the accuracy of pitch detection. Because the speech signal's pre-emphasis is unnecessary, so as to enhance the algorithm's noise-resistance characters.In addition, this paper also did some research on speech signal's formant extraction algorithm, and put forward a formant extraction algorithm based on the Hilbert-Huang Transform. According to the speech's sound channel FM-AM model, using EMD which has adaptive band-pass filter characteristics to separate speech's Formants, estimate the formant's instantaneous frequency and amplitude by energy separation algorithm, according the formant's characteristics of the frequency and amplitude to extract the real formants. The experiment results show that this method not only can accurately extract formant, but also can accurately track the changes of formant.