论文部分内容阅读
：针对各种数字信息，实现了一种基于ＴＭＳ３２０Ｃ５ｘ 评价模块（ＥＶＭ）的与特定人无关的连接数字语音识别系统．在分析了连续概率密度的隐马尔可夫模型（ＣＤＨＭＭ）基础上，利用ＬＰＣ倒谱系数、ＬＰＣ差分倒谱系数、能量归一化系数及其差分系数作为语音特征矢量，训练和识别采用Ｖｉｔｅｒｂｉ算法和Ｂａｕｍ －Ｗｅｌｃｈ 重估算法，有效地提高了系统的识别率．给出了实现各个阶段所需的时间，比较了简单模板匹配法和隐马尔可夫模型法以及不同语音特征参数对识别率的影响．在具体实现中，着重处理了抗噪及实时实现问题．实验结果表明，本系统在普通机房条件下取得较满意的效果，正确识别率达到９２％ ，为其实用化提供了较为重要的技术途径．
: For a variety of digital information, to achieve a based on TMS320C5x evaluation module (EVM) and has nothing to do with the specific connection digital speech recognition system. Based on the analysis of the continuous probability density Hidden Markov Model (CDHMM), the LPC cepstrum coefficients, the LPC difference cepstrum coefficients, the energy normalization coefficients and their difference coefficients are used as the speech eigenvectors. The training and identification are performed using the Viterbi algorithm And Baum-Welch revaluation algorithm, effectively improve the recognition rate of the system. The time required to achieve each phase is given, and the effects of simple template matching and Hidden Markov Models and different speech parameters on the recognition rate are compared. In the concrete realization, it focuses on the anti-noise and real-time implementation. The experimental results show that the system achieves more satisfactory results under the condition of ordinary computer room, and the correct recognition rate reaches 92%, which provides a more important technical approach for its practical application.