基于HMM/ANN混合模型的带噪语音识别

Noise Speech Recognition Based on HMM/ANN

对于含噪语音信号的有效特征提取是语音识别至关重要的一步。该文提出了利用小波调制尺度对语音进行特征提取,结合隐马尔可夫和人工神经网络混合模型进行识别的方法,可进一步反映语音信号的动态特性、增强抗干扰能力和提高识别率。实验证明,该模型适合于对噪声背景下的语音进行识别,同传统的HMM模型相比,具有更好的抗噪鲁棒性,在信噪比较低情况下,识别率比传统的HMM模型有明显的提高。

In speech recognition it is a vital step to extract the effective features of noisy speech signal.This paper presents a new method to extract effective features of noisy speech signal using wavelet modulate scale.Artificial neural network and hidden Markov model were also used to improve the dynamic of speech signal,enhance anti-interference ability and advance speech recognition rate.Experiment showed this model was better than HMM in recognition rate and robustness of noisy signal.