分组网络环境下的实时语音质量客观评价

提出了一种利用前馈随机神经网络在分组网络中进行实时语音质量评价的新方法．从接收到的语音分组中提取美尔频率倒谱系数向量，利用实时传输控制协议计算语音分组传输过程中的丢包率、延迟和抖动，构成网络传输参数向量．将随机神经元组织成具有1个输入层、1个隐含层和1个输出层的3层前馈网络结构，再以上述2种向量作为输入的多类别信号，以相应的主观平均意见（MOS）评分值作为输出对网络进行训练，从而获得稳定的权值矩阵．利用训练过的网络进行多类别信号的语音质量评分映射，并将映射结果与MOS进行二次多项式拟合，得到最终的语音质量评分值．实验表明，所提算法与主观评价之间的平均相关度可达到0．881．

Objective Evaluation for Real-Time Speech Quality in Packet Networks

A novel approach to the real-time speech quality evaluation in a packet network using feed-forward multiple class random neural network(FFMCRNN) was presented.Mel frequency cepstrum coefficient(MFCC) vectors are extracted from the received speech packets.Loss rate,delay and jitter are computed using real-time transport control protocol(RTCP) to form the transport coefficients vector.Random neurons are organized into a 3-layer feed-forward network with one input layer,one hidden layer and one output layer.The network is trained to obtain a stable weight matrix,with the two kinds of vectors above as multi-class input signals and the corresponding subjective mean opinion score(MOS) as output.The multi-class signals are mapped onto the speech quality evaluation scores by using the trained network.The results of mapping are fitted to MOS with a second-order polynomial to get the final speech quality score.Experimental results show that the average correlation between the proposed method and the subjective evaluation was 0.881.