基于子带能量特征的最优化语音端点检测算法研究

Optimization of speech endpoint detection base on sub-band energy feature

为了提高噪声环境下语音端点检测的鲁棒性,提出了一种结合多子带能量特征和最优化边缘检测判决准则的算法。该算法的突出优点在于:在不同信噪比情况下,其端点检测滤波器的输出基本不变,从而避免了门限调整所带来的困难。实验结果表明,这种算法在多种噪声环境下都能够达到较好的语音检出效果。这种算法克服了传统语音端点检测以短时能量、基频、过零率等作为检测特征时,需要动态调整门限且在低信噪比情况下鲁棒性较差的缺点。

In order to detect more robustly and precisely the endpoints of speech under noisy environments, an algorithm was proposed in this paper by combining the multiple sub-bands energy as feature and optimal edge detection as decision criteria. The algorithm highlights itself by exemption of the adjustment of the decision threshold due to the stable output of the filters under the environments with different signal-to-noise ratio. Experiments showed that, while easy to tune the parameters, the algorithm can work more robustly under various noisy environments. Thus overtaking the traditional short-time energy, zero crossing rate and pitch based methods.