公开(公告)号：KR100893154B1

公开(公告)日：2009-04-16

申请号：KR1020080100278

申请日：2008-10-13

Applicant: 한국과학기술연구원 ,  한국전자통신연구원

Inventor： 김형곤 ,  장준혁 ,  강상익 ,  배창석

IPC: G10L15/02

Abstract: A method for recognizing the sex of a voice signal and an apparatus therefor are provided to recognize the sex by differently applying the weighted values which are produced by using the distinctive weight learning according to each degree of an MFCC feature vector, thereby improving the sex recognition performance. A voice signal is inputted(S501). From the inputted voice signal, an MFCC feature vector is extracted(S502). According to each degree of the extracted MFCC feature vector, distinctive weighted values are given(S503). The sex of the inputted signal is recognized by a decision formula to which the distinctive weighted values given according to each of the degrees are applied. Particularly, the sex is distinguished through a value, obtained by multiplying the weighted values with each degree of MFCC feature vector, and a kernel function.

Abstract translation: 提供了一种用于识别语音信号的性别的方法及其装置，以通过不同地应用通过使用根据MFCC特征向量的各个度的独特权重学习产生的加权值来识别性别，由此改善性别识别 性能。 输入语音信号（S501）。 从输入的语音信号中，提取MFCC特征向量（S502）。 根据提取的MFCC特征向量的每个程度，给出特征加权值（S503）。 通过应用根据每个度数给出的特征加权值的判定公式来识别输入信号的性别。 特别地，通过将​​加权值与MFCC特征向量的各个度乘以获得的值和核函数来区分性别。

公开(公告)号：KR100839065B1

公开(公告)日：2008-06-19

申请号：KR1020070076351

申请日：2007-07-30

Applicant: 한국과학기술연구원 ,  한국전자통신연구원

Inventor： 김형곤 ,  장준혁 ,  강상익 ,  조규행

IPC: G10L25/78 ,  G10L25/84 ,  G10L15/04 ,  G10L15/02

Abstract: A voice detector and a voice detection method are provided to apply an optimized weight derived from a minimum classification error scheme to each frequency band likelihood ratio and calculate the geometric average value of the each frequency band likelihood ratio, thereby improving voice detection performance and detecting a voice signal efficiently even in environment with heavy noise. A voice detector comprises an input unit(10) and an analysis unit(20). The input unit receives a signal including a voice signal and performs the DFT(Discrete Fourier Transform) conversion of the received signal. The analysis unit comprises a likelihood ratio calculating module(21), a weight applying module(22), and a voice detecting module(23). The likelihood ratio calculating module calculates a likelihood ratio by frequency band of the converted signal. The weight applying module multiplies the likelihood ratio by frequency band by different weights to calculate the likelihood ratio to which the weight is applied. The voice detecting module compares a geometric average value of the calculated likelihood ratios with a preset value to measure whether the converted signal includes the voice signal. Further, the weight applying module calculates weight in frequency band by an MCE(minimum classification error) method.

Abstract translation: 提供语音检测器和语音检测方法，对每个频带似然比应用从最小分类误差方案导出的优化权重，并计算每个频带似然比的几何平均值，从而提高语音检测性能并检测 即使在环境噪音较大的情况下也能有效地发出语音信号。 语音检测器包括输入单元（10）和分析单元（20）。 输入单元接收包括语音信号的信号，并执行接收信号的DFT（离散傅里叶变换）转换。 分析单元包括似然比计算模块（21），加权应用模块（22）和语音检测模块（23）。 似然比计算模块根据转换后的信号的频带计算似然比。 权重应用模块将似然比乘以频带不同的权重，以计算应用权重的似然比。 语音检测模块将计算出的似然比的几何平均值与预设值进行比较，以测量转换的信号是否包括语音信号。 此外，权重施加模块通过MCE（最小分类误差）方法计算频带的权重。