Towards the automatic processing of Yongning Na (sino-tibetan): developing a 'light' acoustic model of the target language and testing 'heavyweight' models from five national languages.