公开(公告)号：KR100269254B1

公开(公告)日：2000-10-16

申请号：KR1019970072163

申请日：1997-12-23

Applicant: 한국전자통신연구원

Inventor： 한문성 ,  김정환 ,  황두성 ,  박전규 ,  박재득 ,  박동인 ,  박규봉

IPC: G10L13/00

Abstract: PURPOSE: A method for recognizing using a predictive neural network is provided to improve of an existing LPNN(Linked Predictive Neural Network) based voice recognition algorithm by using a voice recognition method. CONSTITUTION: Input voice features are normalized to voice frames(F1-F18). Each corresponding frame(F1-F18) are connected with predictive neural networks(P1-P16) of the fixed number. A predictive process is performed. A predictive value is outputted through the corresponding predictive neural network. A practical value is outputted from a corresponding frame. An accumulated predictive error is outputted by comparing the predictive value with the practical value. The accumulated predictive error is used as a reference of decision of voice recognition.

Abstract translation: 目的：提供一种使用预测神经网络进行识别的方法，通过使用语音识别方法改进现有的基于LPNN（基于预测神经网络的链接预测神经网络）语音识别算法。 规定：输入语音功能对语音帧进行归一化（F1-F18）。 每个对应的帧（F1-F18）与固定数量的预测神经网络（P1-P16）连接。 执行预测过程。 通过相应的预测神经网络输出预测值。 从对应的帧输出实用值。 通过将预测值与实际值进行比较来输出积累的预测误差。 累积的预测误差被用作语音识别决策的参考。

公开(公告)号：KR1019980004336A

公开(公告)日：1998-03-30

申请号：KR1019970072163

申请日：1997-12-23

Applicant: 한국전자통신연구원

Inventor： 한문성 ,  김정환 ,  황두성 ,  박전규 ,  박재득 ,  박동인 ,  박규봉

IPC: G10L13/00