Speaker recognition feature extraction method based on PSNCC (perception spectrogram Norm cochlea-filter coefficient)

Abstract

The invention discloses a speaker recognition feature extraction method based on a PSNCC (perception spectrogram norm cochlea-filter coefficient). The speaker recognition feature extraction method comprises the following steps: first, constructing a cochlea filter group conforming to traveling wave impulse response and nolinear frequency distribution of a cochlear basilar membrane; then carrying out speech enhancement and two-dimensional enhancement based on auditory perception features on voice, and carrying out two-dimensional boundary detection on a pure voice spectrum structure in continuous distribution, thus obtaining a perception spectrogram structure norm parameter PSN; finally, further normalizing all cochlea-filter coefficients output by the cochlea filter group in the time domain through the perception spectrogram structure norm parameter PSN, and extracting the PSNCC feature parameter. According to the PSNCC feature parameter extracted by adopting the method, the robust performance of the feature parameter is improved from two aspects, namely, the time domain and the frequency domain, and further, the recognition rate of a speaker recognition system under the noise environment with low signal to noise ratio is improved.

Description

technical field [0001] The invention relates to the technical field of speech recognition, in particular to a method for extracting speaker recognition features by perceiving spectrum regularized cochlear filter coefficients. Background technique [0002] Feature extraction is the first important component in speaker recognition. In general, successful front-end feature parameters should carry enough judgmental information for classification or recognition, be suitable for back-end modeling, and be fairly robust to changes in the acoustic environment. However, achieving satisfactory system performance under different operating modes remains problematic, especially when the auditory training and testing environments are severely mismatched. The robust performance of feature parameters in low SNR environment needs to be further improved to solve this mismatch. [0003] One of the most common features is the Linear Prediction Cepstral Coefficients (LPCC). The speech formant ...

AI Technical Summary

Problems solved by technology

[0008] Technical problem to be solved: In order to solve the problem that the robustness of the traditional feature parameter processing method in the frequency domain is not strong, the recognition rate of speaker recognition drops sharply when the signal-to-noise ratio decreases, and the current feature parameters pay more attention to the frequency domain processing to obtain improved robustness, ignoring that the feature extraction process is a complete time-frequency domain process, a speaker recognition feature extraction method that perceives regularized cochlear filter coefficients is proposed

Images