基于日常会话的司法话者非语音学识别特征挖掘研究

A Study of Non-phonetic Features in Forensic Speaker Recognition Based on Daily Conversations

目的目前司法话者识别特征以语音学特征为主,语音学特征的不足是其受现实环境因素影响大,产生语音变异现象,导致鉴定结果信度降低。挖掘可用于司法话者识别的非语音学识别特征,以达到提高鉴定结果信度的目的。方法采用不控制任何现实环境因素和话者生理与心理因素条件下采集的自然日常会话为语料,依据Sapir对话语行为结构层面的描述和认知语言学"现实—认知—语言"这一基本思想,采用语篇信息分析方法挖掘具有潜在话者识别能力的个人话语风格特征,并设计实验验证其作为话者识别特征的性能、效度和信度。结果实验结果显示,挖掘的个人话语风格特征具有高话者间差异性和低话者内差异性,而且其识别能力受现实环境因素和话者个人因素影响都很小。结论挖掘的个人话语风格特征可以作为潜在的话者识别特征。在司法实践中如果非语音学的个人话语风格特征能与话语行为其他层面上的高效度识别特征,尤其是语音学特征,互为补充,识别结果相互印证,识别结论的信度会大大提高。

Objective Phonetic features are the main type of forensic speaker recognition features. One problem with phonetic features is that they are affected dramatically by the real-world conditions, which results in intra-speaker variation and consequently reduces the reliability of forensic speaker cognition results. This study aims to explore non-phonetic features that can be applied to forensic speaker recognition to improve the reliability of forensic speaker cognition results. Method Supported by Sapir's description of the structure of speech behavior and the basic idea of "reality-cognition-language" from cognitive linguistics, individual speaking style features were extracted by applying discourse information analysis to daily conversations. Experiments were designed to test their general performance, validity and reliability, as potential forensic speaker recognition features. Results Individual speaking style features extracted from daily conversations showed high inter-speaker variation and low intra-speaker variation, and were less affected by real-world conditions. Conclusion The individual speaking style features explored in this study are qualified to be potential forensic speaker recognition features. The reliability of forensic speaker recognition results could be improved if non-phonetic features are employed as the complement to highly valid phonetic features for cross-validation.