本文为美国马凯特大学（作者：Anthony D. Ricke）的硕士论文，共119页。

自动语音识别（ASR）是一种有助于研究动物发声的有用工具，在动物发声中使用基于人类语言的信号处理技术具有多种缺陷。动物发声可能与人类语言不具有相同的频谱或时间特征。因此，关于最佳帧长度、帧重叠和HMM拓扑结构的典型ASR假设可能不适用于各种动物发声。

本文提出了一种从纯净的动物发声实例中估计帧长度、帧重叠和HMM拓扑结构的技术。针对两种不同动物的发声进行了研究：挪威圃鹀（Emberiza Hortulana）和非洲象（Loxodonta Africana），采用所提出的技术进行了多次试验。通过对试验结果的分析，给出了该算法的合理估计，并给出了该算法的帧长、帧重叠和HMM拓扑结构。为将本研究作为动物发音研究的有用工具继续下去，提出了具体的建议。

Automatic Speech Recognition (ASR) is auseful tool that can facilitate the research and study of animal vocalizations.The use of human speech-based signal processing techniques for animalvocalizations has several pitfalls. Animal vocalizations may not share the samespectral or temporal characteristics as human speech. As a result, the typicalASR assumptions concerning the best frame length, frame overlap and HMMtopology may not be suitable for various animal vocalizations. This paperproposes a technique for estimating the frame length, frame overlap and HMMtopology from a single, clean, example animal vocalization. Multiple trials arerun using the proposed technique, against the vocalizations of two distinctanimal species: the Norwegian Ortolan Bunting (Emberiza Hortulana) and the AfricanElephant (Loxodonta Africana). The results are examined, and the techniqueprovides reasonable estimates for the frame length, the frame overlap and theHMM topology, given the quality of the example vocalizations. Specificrecommendations are made for the continuation of this research into a usabletool for animal researches.

1 引言

2 项目背景与相关工作

3 本文所提出的研究方法

4 总结

附录A 软件设计

更多精彩文章请关注公众号：