基于时频结合的带噪语音端点检测算法

Endpoint detection algorithm for noisy speech based on time-frequency combination

为了改善在低信噪比条件下,传统语音端点检测算法准确率较低的情况,提出了一种结合多窗谱估计的谱减法和能熵比的语音端点检测算法。该算法在低信噪比下,对带噪语音进行多窗谱估计的谱减语音增强后,结合语音信号的短时能量和子带谱熵,对增强后的语音信号进行能熵比的计算,并用于端点检测。实验结果表明,在不同的背景噪声且信噪比为-5 d B环境下,相对其他端点检测算法更有效地检测出语音端点,可达到70%以上的正确率,此算法更适合于低信噪比环境下的语音端点检测。

In order to improve the low accuracy of the traditional speech endpoint detection in low SNR condition, this paper proposes a speech endpoint detection algorithm based on the combination of spec- trum subtraction with muhitaper spectrum estimation and energy-entropy ratio. Muhitaper spectrum esti- mation of spectrum subtraction speech enhancement is performed at the first step. Combining with the en- ergy and the sub-band spectral entropy, the energy-entropy ratio is obtained to be used as the endpoint detection parameter. Experimental results show that, in a low SNR = -5 dB with different background noises, the algorithm proposed can operate more effectively than other extant endpoint detection algo- rithms. With this algorithm, the detection accuracy can achieve 70% or above, which further proves that the proposed algorithm is suitable for noisy speech endpoint detection in low SNR environment.