A Siamese Neural Network Based Method for Lineage Determination of Speech Samples for Forensic Applications

The research is designed to build a Siamese Neural Network to determine the lineal relationship between speech samples. For this research, the TIMIT Acoustic-Phonetics continuous speech corpus was obtained under a data scholarship from Linguistic Data Consortium. The speech samples in TIMIT dataset where further modified using mathematical functions, and processed for Mel Frequency Cepstral Coefficients (MFCCs) extraction. Hence, our dataset consists of normal and frequency distorted dataset.

本研究旨在构建孪生神经网络（Siamese Neural Network），以判定语音样本间的线性关联关系。本研究通过语言数据联盟（Linguistic Data Consortium, LDC）的数据资助计划，获取了TIMIT声学-语音连续语音语料库。TIMIT数据集中的语音样本经数学函数进一步修改，并经处理以提取梅尔频率倒谱系数（Mel Frequency Cepstral Coefficients, MFCCs）。因此，本数据集包含正常语音样本与经过频率畸变处理的语音样本两类。