语音识别SoC UniLite的系统设计

UniLite系统以DSP为核心，集成了片内存储器、AD/DA通道和多种的输入输出端口，有效控制了系统的成本和功耗，在有限的RAM空间上实现了高性能非特定人语音识别以及语音编解码的功能。系统采用连续隐含马尔科夫（CHMM）算法，运用了分阶段提取特征、束搜索等策略，在保证系统识别性能的同时大大降低了内存消耗，提高了识别速度，识别率在98.5%以上，识别时间在0.5倍实时 以下。

System Design of Speech Recognition SoC UniLite

The UniLite chip is realized around an architecture which is composed of a 16-bit DSP core, on-chip memory, an ADC/DAC channel as well as other common I/O interfaces, High performance speaker independent speech recognition and speech coding are implemented on limited RAM memory. This results in low power consumption and reduction in cost. The system is based on continuous hidden Markov model. Two-pass feature extraction and beam search strategy are also employed to reduce the consumption of the RAM and enhance the recognition rate. Test shows that the UniLite system provides recognition accurate as high as 98.5% and requires less than 0,5 real time.