This paper analysed and contrasted multi-algorithms about DHMM (Discrete Hidden Markov Models), based on which it designed an improved DHMM algorithm, and then, applied it to the embedded speech recognition system. This system was realization of speaker-independent’s isolated word recognition. The flatform of this system was centered ADSP-BF531. And it was expanded by big capacity SDRAM & ROM. This system has many merits, e.g. small volume, high speedand, high reliability, high robustness and high expansibility, etc. It is very convenient for some special situations and has much market potential. The experiment confirms that its speech recognition accuracy reaches 94 percent for speaker-independent and small vocabulary. This paper introduced the improved HMM algorithm, hardware design and its application effect.