基于DTW的俄语短指令语音识别

发布时间：2018-01-10 07:02

本文关键词：基于DTW的俄语短指令语音识别　出处：《山东大学学报(理学版)》2017年11期 　论文类型：期刊论文

【摘要】：面向训练语料有限的语音识别任务,基于动态时间规整(dynamic time warping,DTW)算法对俄语语音进行识别。首先,以跨语言标注的语音语料为资源基础,研究融合音字转换和机器翻译的语音识别方法。其次,结合俄语语音特点,以元音为中心设置动态门限阈值,实现精确至音节的端点检测,识别速度提高了34.4%,准确率提高了14%。然后,综合时域、频域分析,提取反映语音静态特征和动态变化的参数模板。另外,引入全局限制和早弃策略改进DTW算法,避免病态匹配,缩小计算规模,使速度提高了19.7%,准确率提高了4.8%。在俄语短指令语音集上做五折交叉验证,识别准确率达到74.9%。
[Abstract]:For the task of speech recognition with limited training data, Russian speech recognition is carried out based on dynamic time warping (DTW) algorithm. Based on the cross-language tagged speech corpus, this paper studies the speech recognition method combining phonetic word conversion and machine translation. Secondly, combining the Russian speech characteristics, the vowel-centered dynamic threshold is set up. To achieve accurate to syllable endpoint detection, recognition speed increased 34. 4, accuracy improved 14. Then, integrated time domain, frequency domain analysis. In addition, the global restriction and early abandonment strategies are introduced to improve the DTW algorithm to avoid ill-posed matching, reduce the scale of calculation, and improve the speed by 19.7%. The accuracy is improved by 4.8. the recognition accuracy is 74.9% by 50% cross-verification on the Russian short instruction speech set.
【作者单位】：中国人民解放军外国语学院语言工程系;
【基金】：国家自然科学基金重大项目(11590771)
【分类号】：TN912.34
【正文快照】： 0引言语音识别(automatic speech recognition,ASR)是指机器通过识别和理解过程将语音信号转换为相应的文本或命令[1],然后根据信息内容,执行人的某种意图[2]。面向汉语和英语的大词汇连续语音识别在统计建模、神经网络等机器学习方法的推动日趋实用化。但针对小语种、少数

【相似文献】