User Friendly Speech Recognition Algorithm with Adaptability for Environments and Users

用户友好的语音识别算法，具有对环境和用户的适应性

基本信息

批准号：
15300060
负责人：
SHIKANO Kiyohiro
金额：
$ 10.3万
依托单位：
Nara Institute of Science and Technology
依托单位国家：
日本
项目类别：
Grant-in-Aid for Scientific Research (B)
财政年份：
2003
资助国家：
日本
起止时间：
2003 至 2006
项目状态：
已结题

项目摘要

The research plan includes the following topics,1. Noise reduction signal processing, accurate phoneme modeling, and speaker and environment adaptation.2. Task adapted language model for accepting spontaneous utterances.3. Hands-free speech recognition interface.4. Study of human factors for speech dialog system.These investigations have been studied based on real world speech dialog systems.The main attained research results are summarized as follows.1. Unsupervised speaker adaptation algorithm based on an arbitrary utterance has been developed, which takes only several seconds, and shows almost same accuracy as supervised adaptation MLLR based on 50 utterances.2. Task adapted language models have been developed for children and adults using two year Takemaru-kun transcribed texts. By introducing parallel decoding for children and adults, we attain the improvements of word accuracy and response accuracy.3. Hands-free speech recognition has been implemented based on null-beamformer type SSA (Spatial Subtraction Array) and BSSA (Blind SSA) with SIMO-ICA.4. We have been successfully operating Takemaru-kun speech guidance system in Ikoma North community center these four years. We have been also operating two speech guidance systems in local railway station this one year, to study about noisy condition. The collected speech database with the transcription is useful to develop speech dialog systems.We also invented a new quiet speech media, Non-Audible Murmur (NAM), which applied to quiet speech recognition and quiet telephone. This invention awarded by IEICE paper prize and Inose award.

研究计划包括以下主题，1。降噪信号处理，准确的音素建模以及扬声器和环境适应。2。任务改编的语言模型，用于接受自发话语3。免提语音识别接口4。根据现实世界语音对话系统进行了研究的人为因素的研究。这些研究的研究结果总结了如下。1。已经开发了基于任意话语的无监督的说话者适应算法，仅需几秒钟，并且基于50个话语显示与受监督的适应性MLLR几乎相同的准确性。2。使用两年的takemaru-kun抄录文本为儿童和成人开发了任务调整的语言模型。通过为儿童和成人引入并行解码，我们可以提高单词准确性和响应准确性3。无提机语音识别是基于Null-BeamFormer型SSA（空间减法阵列）和BSSA（盲人SSA）实现的。4。这四年，我们已经成功地在Ikoma North社区中心运营了Takemaru-Kun语音指导系统。这一年我们还在当地火车站运营两个语音指导系统，以研究嘈杂的状况。带有转录的收集的语音数据库对于开发语音对话系统很有用。我们还发明了一种新的安静的语音媒体，不可察觉的杂音（NAM），该媒体适用于安静的语音识别和安静的电话。这项发明由IEICE纸奖和Inose奖授予。

项目成果

期刊论文数量（420）

专著数量（0）

科研奖励数量（0）

会议论文数量（0）

专利数量（0）

High-Fildeity Blind Separation of Acoustic Signals using SIMO-Model-Based ICA with Information-Geometric Learning

使用基于 SIMO 模型的 ICA 和信息几何学习对声学信号进行高保真度盲分离

DOI：
发表时间：
2003
期刊：
Proceedings of IWAENC2003
影响因子：
0
作者：
Tomoya Takatani;Tsuyoki Nishikawa;Hiroshi Saruwatari;Kiyohiro Shikano
通讯作者：
Kiyohiro Shikano

Stable Learning Algorithm for Blind Separation of Temporally Correlated Signals Combining Multistage ICA and Linear Prediction

结合多级 ICA 和线性预测的时间相关信号盲分离的稳定学习算法

DOI：
发表时间：
2003
期刊：
Proceedings of Fourth International Symposium on Independent Component Analysis and Blind Signal Separation
影响因子：
0
作者：
Tsuyoki Nishikawa;Hiroshi Saruwatari;Kiyohiro Shikano
通讯作者：
Kiyohiro Shikano

Two-Stage Blind Source Separation Using SIMO-ICA and Binary Masking

使用 SIMO-ICA 和二进制掩蔽的两级盲源分离

DOI：
发表时间：
2005
期刊：
Proceedings of Joint Workshop on Hands-Free Speech Communication and Microphone Arrays (HSCMA2005)
影响因子：
0
作者：
Hiroshi Saruwatari;Yoshimitsu Mori;Tomoya Takatani;Satoshi Ukai;Kiyohiro Shikano;Takashi Hiekata;Takashi Morita
通讯作者：
Takashi Morita

Model-Integration Rapid Training based on Maximum Likelihood for Speech Recognition

基于最大似然的模型集成快速训练语音识别