论文部分内容阅读
在语音识别领域中,我们把线性预测编码(Linear Predictive Coding,缩略简称为LPC)技术的特征检测单片处理机实现称之为“特征提取数字信号处理机(Feature Extracting Digital Singal Processor,缩略简称为FXDSP)。这种实现是在AT&T的DSP20~(TM)可编程数字信号处理器(Digital Singal Processor,缩略简称为DSP)的基础上进行编程而开发出来的,并且已经经过数字仿真和系统运行的检验。输入同样的信号,将浮点仿真与DSP实现之间的识别距离同单词的匹配距离进行比较,发现其识别距离小到可以忽略不计。在LPC技术的数字仿真中,孤立单词及关联单词的识别采用了动态时间扭曲法,向量量化和隐式马尔柯夫模型,其特征检测技术是相同的。AT&T贝尔实验室研究工作的结果发现,FXDSP是一种单片式的公共积木模块,它代表了大多数语音识别实时实现的技术。FXDSP把标准编码译码器中收到的语音进行8阶LPC分析。每一个帧周期(15ms)中产生一个特征向量,该特征向量包括了对数能量、9个振幅规范化的自相关系数以及9个基于LPC检测模式的系数。特征检测程序需在1024个可利用的在电程序ROM中用去1023个存储单元,以及256个可利用的RAM存储单元中用去211个,在可利用的真实时间中用去75％。
In the field of speech recognition, we implement the feature detection single chip processor of Linear Predictive Coding (abbreviated as LPC) technology as a so-called Feature Extracting Digital Signal Processor (Feature Extracting Digital Singal Processor Referred to as FXDSP.) This implementation is based on AT & T’s DSP20 ~ (TM) programmable digital signal processor (Digital Singal Processor, abbreviated as DSP) based on the programming and developed, and has been digitally simulated and The same as the signal, the recognition distance between floating-point simulation and DSP implementation is compared with the word matching distance and found that the recognition distance is small to negligible.In the digital simulation of LPC technology, isolated words And related words recognition using dynamic time warping, vector quantization and implicit Markov model, the characteristics of the detection technology is the same.AT & T Bell Laboratory research found that FXDSP is a single piece of public building blocks Module, which represents most of the real-time implementation of speech recognition technology.FXDSP the standard codec received speech 8-order LP C analysis. A feature vector is generated in each frame period (15ms) that includes logarithmic energy, nine amplitude-normalized autocorrelation coefficients, and nine coefficients based on the LPC detection mode. The signature detection procedure requires 1024 Available in the electronic program ROM to use 1023 storage units, and 256 available RAM storage unit to 211, 75% of the available real time.