Harrington J
Speech Hearing and Language Research Centre, Macquarie University, Sydney, Australia.
J Acoust Soc Am. 1994 Jul;96(1):19-32. doi: 10.1121/1.410465.
Recent studies have shown that the acoustic relationship between the murmur and the vowel at the nasal-vowel boundary is highly informative for the [m]-[n] distinction. In the present paper, the contribution of relational information is reassessed by classifying 1946 syllable-initial and 2848 syllable-final nasal consonants taken from continuous speech data. Relational information in the acoustic waveform is based on difference spectra, in which spectral information in the vowel is subtracted from spectral information in the murmur, and on combined spectra in which classifications are made from combinations of murmur and vowel spectra. These two kinds of relational spectra are compared with static spectra, in which single spectral slices are taken in either the murmur or the vowel. Contrary to recent theoretical predictions, difference spectra are shown to perform more poorly than some kinds of static spectra. However, since classification scores from combined spectra are better than from either static or difference spectra, cues to nasal place of articulation can nevertheless be defined as relational. In the best scoring combined spectra, classification scores on open tests are just under 94% correct for syllable-initial nasals and just under 82% correct for syllable-final nasals. The high classification scores show that there is considerable information in the acoustic waveform for identifying nasal place of articulation from continuous speech data.
最近的研究表明，鼻音与元音边界处的元音之间的声学关系对于区分[m]和[n]具有很高的信息量。在本文中，通过对从连续语音数据中提取的1946个音节开头和2848个音节结尾的鼻音进行分类，重新评估了关系信息的贡献。声学波形中的关系信息基于差分谱，其中从鼻音中的频谱信息中减去元音中的频谱信息，以及基于组合谱，其中分类是根据鼻音和元音频谱的组合进行的。将这两种关系谱与静态谱进行比较，静态谱是在鼻音或元音中获取单个频谱切片。与最近的理论预测相反，差分谱的表现比某些类型的静态谱更差。然而，由于组合谱的分类分数优于静态谱或差分谱，因此仍然可以将鼻音发音部位的线索定义为关系性的。在得分最高的组合谱中，开放测试中音节开头鼻音的分类正确率略低于94%，音节结尾鼻音的分类正确率略低于82%。高分类分数表明，从连续语音数据中识别鼻音发音部位的声学波形中存在大量信息。
