基于STM32单片机的嵌入式语音识别系统设计

设计了一款以STM32F103为核心的自然语言识别系统,为满足实时语音识别系统对内存资源和运算速度的要求,基于硬件资源合理设计语音处理算法,在嵌入式平台上实现了对孤立词语的语音识别。首先根据背景噪声和语音信号的时域特征差异设定相应门限值,从而实现了对语音信号的端点检测。然后针对语音识别中传统梅尔倒谱系数对语音的高频信息敏感度较低,对语音信号分别提取梅尔倒谱系数(MFCC)与翻转梅尔倒谱系数(IMFCC),结合Fisher准则构造混合特征参数。最后采用动态时间规整算法实现语音识别。因系统体积小、便携性好等特点,易于实现对不同设备的语音控制,有一定的市场前景。

A natural language recognition system is designed based on STM32F103.To meet the requirements of real-time speech recognition system for memory resources and computing speed,the speech processing algorithm is designed based on hardware resources and speech recognition of isolated words is implemented on the embedded platform.Firstly,the corresponding threshold is set according to the time domain characteristic difference of the speech signal and the background noise and thereby realizing the endpoint detection of the speech signal.Concerning the traditional Mel Frequency Cepstral Coefficient(MFCC)in speech recognition is less sensitive to high frequency signals of speech,MFCC and IMFCC(Inverted MFCC)are extracted respectively for the speech signal and the Fisher criterion is used to construct the mixed feature parameters.Dynamic time warping algorithm is used in speech recognition process.Due to the small size of the system and good portability,it is easy to implement voice control for different devices and has much marker potential.

点击标签浏览更多相关技术资料资源：