基于功率谱和共振峰的母羊发声信号识别

Vocal signal recognition of ewes based on power spectrum and formant analysis method

内蒙古及周边西部地区正在发展规模化种草设施圈养,这种养殖模式要求较高的福利化饲养水平。母羊在不同的应激行为下会发出不同的声信号,可以通过识别母羊发声信号去评价其健康状况和福利化养殖水平。该研究以成年小尾寒羊为例,通过无线语音数据采集卡,平均采集80只母羊在寻羔、饥饿和惊吓3种应激行为下的发声,用Audacity软件共分割成1 200句叫声信号,并用带通滤波和小波消噪进行预处理。每种应激行为下再随机选取200句发声信号,共计600句进行AR(auto-regressive)功率谱估计和共振峰分析,提取第1、2和3共振峰频率和6个代表性的功率谱估计频域参数:功率谱密度的平均值、几何平均值、中值、切尾平均值、平均绝对偏差值和四位分极差,同时也提取叫声信号的最大值、持续时间和间隔时间时域参数,这些特征参数用于训练BP(back propagation)神经网络母羊发声信号识别模型,剩余的600句发声信号用于测试模型的识别效果。结果表明:母羊在不同应激行为下的发声信号具有明显差异的特征参数,采用共振峰参数训练的BP网络,其对母羊发声信号的正确识别率为85.3%,高于利用AR功率谱估计参数的81.0%,当2种参数进行组合训练BP网络后,其正确识别率可达93.8%,表明这种方法的识别效果更好,由于在同一种应激行为下,不同年龄和体质量的母羊发声信号具有一定的差异性,使得系统的误识别率达到6.2%。

Inner Mongolia and its surrounding areas in the west are developing an intensive and large-scale sheep farming operation, in which sheep are bred with planting forage and are placed in captive facilities. However, the breeding pattern of such operation needs a high level of animal welfare management. Considering that sheep makes different vocalization in different emergent situations, ewes’ vocalization can be used as an important evidence for ewes’ health monitoring and breeding welfare evaluation. In this paper, taking Small Tail Han sheep as an example, ewes’ vocalization signals were evenly collected from 80 adult ewes under 3 stress behaviors including searching lamb, hunger, and scare via a wireless audio surveillance device. Then, these continuous vocal signals of ewes were split into 1 200 single call signals using Audacity Acoustic Edit software. The band-pass filter and wavelet denoising methods were applied to preprocess those single sound signals. Six hundred of those sound signals, which were comprised of three different stress behaviors by random selected 200 signals, were analyzed to extract ewes’ acoustic characteristic parameters using auto-regressive (AR) power spectrum estimation and formant extraction methods, respectively. Therefore, six representative frequency characteristic parameters from AR power spectrum estimation method were extracted: the power spectrum density mean, the geometric mean, the median value, the trimmed mean, the mean absolute deviation, and inter quartile deviation, and characteristics parameters from formant analysis method were the first, second and third formant frequency. Moreover, typical time-domain characteristic parameters such as signal maximum value, duration value and interval value were taken as well. Then, these characteristic parameters were used to train the back propagation (BP) neural network model of ewes’ vocalization recognition, and the rest of 600 vocal signals were used to test the effects of the recognition mode. The results demonstrated that characteristic parameters of ewes’ vocal signals were obviously different under different stress behaviors. Furthermore, if BP recognition network was trained by formant parameters, the average correct recognition rate of ewes’ vocal signal was 85.3%, higher than AR power spectrum estimation parameters of 81.0%. When BP network was trained by a combination of above two kinds of characteristic parameters, the average correct recognition rate was 93.8%, which meant that the performance of the combination parameters was better than another two methods. However, the average false positive rate still reached 6.2% because ewes’ vocal signals under the same stress behavior had a certain degree of difference due to the different age and weight as well as sound volume strength. The results of this study also indicated that analysis of vocalization could be an indicator of different physiological conditions in sheep and may be an important role for understanding communications in ewes.