有色噪声环境中鲁棒语音特征参数提取研究

摘要

针对复杂噪声干扰环境中语音特征参数会发生改变,引起训练模型和测试语音之间的失配,使语音识别系统的识别率降低,为提高语音特征参数在色噪声环境中提取的鲁棒性,提出了基于总体最小二乘旋转不变子空间技术(TLS-ESPRIT)谐波倒谱加权谱鲁棒特征参数提取方法.运用TLS-SVD方法对观测数据矩阵进行广义特征值分解估计谐波模型的参数,实现了有色噪声背景下语音信号的最优估计.在重建语音的过程中根据谐波能量与带噪语音能量的比值,对重建谐波的各个谐波峰给予不同的加权和语音建模,并进行仿真,结果实现了鲁棒性特征参数的提取,解决了模型之间的失配问题.%Speech feature parameters would change in Complex noise environment, causing the mismatch between training model and test speech.The recognition rate of speech recognition system would be reduced.In order to enhance the robustness of speech features extraction in the colored noise environment, a method of harmonic weighted spectrum of robust cepstral feature extraction was proposed based on Total Least Square Estimation of Signal Parameters Via Rotational Invariance Techniques.The parameters of harmonic model could be estimated by using the TLSVAD algorithm to calculate the generalized eigenvalue decomposition from the observation data matrix, to achieve optimal estimation of speech signals under colored noise background.The different weights of harmonic peaks of each frame were calculated according the ratio of the harmonic frame energy with the noisy speech energy in reconstruction process to achieve a robust feature parameter extraction.