论文部分内容阅读
为了达到嵌入式语音识别系统低成本、低功耗的目标,提出一种算法硬件映射方法。将基于连续隐含Markov模型语音识别算法中占系统总运算量的50%以上的Mahalanobis距离计算,映射为硬件实现的模块。通过该方法,系统在较低时钟频率下即可完成嵌入式语音识别中实时处理的要求,从而大大降低系统功耗。实验结果表明,该模块在0.18μm和舰工艺库下实现,仅需1.2mm2,包含64kb静态随机存储器。应用该模块可以大大提高嵌入式语音识别系统的性能,达到降低成本,降低功耗的目标。
In order to achieve the goal of low cost and low power consumption of embedded speech recognition system, an algorithm hardware mapping method is proposed. The Mahalanobis distance calculation based on continuous implicit Markov model speech recognition algorithm accounting for more than 50% of the total system operation is mapped to a hardware implemented module. By this method, the system realizes the requirement of real-time processing in embedded speech recognition at a lower clock frequency, thereby greatly reducing system power consumption. The experimental results show that the module is implemented in 0.18μm and ship technology library, only 1.2mm2, including 64kb static random access memory. Application of the module can greatly improve the performance of embedded speech recognition system, to reduce costs and reduce power consumption goals.