A Method of Dynamic Feature Extraction of Speech Signal Based on Formant Curve

Abstract

The invention proposes a method for extracting dynamic features of speech signals based on formant curves, which belongs to the technical field of extracting dynamic features of Chinese speech signals. The steps are: collecting the speech signal; preprocessing the speech signal; extracting the formant frequency feature of the speech signal; according to the frame sequence from the first frame to the last frame, the first formant frequency Combine the eigenvalues ​​to obtain the first formant curve, and so on, to obtain the second formant curve, the third formant curve and the fourth formant curve; fast Fourier transform is performed on each obtained formant curve to obtain linearity Spectrum; obtain energy spectrum from linear spectrum; obtain logarithmic energy from energy spectrum; perform discrete cosine transform on logarithmic energy. Compared with the existing method, the present invention extracts the dynamic feature of the speech signal, which has time correlation, reveals the close correlation between the front and back of the speech signal and between adjacent speech signals, and improves the performance of speech recognition.

Description

technical field [0001] The invention belongs to the technical field of dynamic feature extraction of Chinese speech signals, in particular to a method for extracting dynamic features of speech signals based on formant curves. Background technique [0002] Speech recognition research in my country started in the 1950s, but it did not develop rapidly until the 1970s. The Chinese Academy of Sciences, Tsinghua University, Peking University and many other research institutes are engaged in the development of Chinese speech recognition system. At present, the research on continuous speech recognition system with large vocabulary is close to the highest level in foreign countries; "In the plan, the research on Chinese speech recognition has received strong support. The National 863 "Intelligent Computer Topics" expert group has specifically established a project for speech recognition research. At the same time, due to China's growing international status and its important position...

AI Technical Summary

Problems solved by technology

With the development of recognition technology, people found that the characteristic parameters in the time domain are not very stable and distinguishable, so they began to use the frequency domain parameters as the characteristics of the speech signal, such as pitch period, formant frequency, linear prediction coefficient ( LPC), line spectrum pair (LSP), cepstral coefficient, etc., the most widely used feature parameter is the Mel cepstral coefficient (MFCC) based on the human auditory model; but once these parameters are applied to the noise environment, their performance will drop sharply;

[0005] Moreover, the characteristic parameters mentioned above all reflect the static characteristics of the speech. The dynamic characteristics of the speech signal refer to the characteristic parameters extracted from several consecutive frames of speech. Acceleration parameters cannot fully mine the dynamic information, so they cannot reflect the dynamic characteristics of the speech signal well

Images