School of Music, University of Edinburgh, City of Edinburgh EH9 3JZ, United Kingdom.
J Acoust Soc Am. 2012 Jun;131(6):4785-98. doi: 10.1121/1.4707535.
Physiological evidence suggests that sound onset detection in the auditory system may be performed by specialized neurons as early as the cochlear nucleus. Psychoacoustic evidence shows that the sound onset can be important for the recognition of musical sounds. Here the sound onset is used in isolation to form tone descriptors for a musical instrument classification task. The task involves 2085 isolated musical tones from the McGill dataset across five instrument categories. A neurally inspired tone descriptor is created using a model of the auditory system's response to sound onset. A gammatone filterbank and spiking onset detectors, built from dynamic synapses and leaky integrate-and-fire neurons, create parallel spike trains that emphasize the sound onset. These are coded as a descriptor called the onset fingerprint. Classification uses a time-domain neural network, the echo state network. Reference strategies, based upon mel-frequency cepstral coefficients, evaluated either over the whole tone or only during the sound onset, provide context to the method. Classification success rates for the neurally-inspired method are around 75%. The cepstral methods perform between 73% and 76%. Further testing with tones from the Iowa MIS collection shows that the neurally inspired method is considerably more robust when tested with data from an unrelated dataset.
生理证据表明，听觉系统中的声音起始检测可能早在耳蜗核就由专门的神经元完成。心理声学证据表明，声音起始对于音乐声音的识别很重要。在这里，声音起始被单独用于为乐器分类任务形成音调描述符。该任务涉及来自 McGill 数据集的五个乐器类别的 2085 个孤立音乐音。使用受听觉系统对声音起始反应启发的模型创建神经启发式音调描述符。伽马滤波器组和尖峰起始检测器，由动态突触和漏电流积分和放电神经元构建，创建强调声音起始的并行尖峰序列。这些被编码为称为起始指纹的描述符。分类使用时域神经网络，回声状态网络。基于梅尔频率倒谱系数的参考策略，无论是在整个音上还是仅在声音起始期间进行评估，都为该方法提供了上下文。神经启发方法的分类成功率约为 75%。基于倒谱的方法的性能在 73%到 76%之间。使用爱荷华 MIS 集合中的音进一步测试表明，当使用来自不相关数据集的数据进行测试时，神经启发方法的稳健性大大提高。
