||Many practical speech recognition applications such as in information retrieval need to be portable to lightweight client devices such as PDAs. We propose that a two-stage recognizer with low memory requirement in the first stage is desirable. It is well known that Chinese language is syllabic in nature. All the Chinese words consist of one to several Chinese characters and all these Chinese characters are monosyllabic. Therefore, we can perform Chinese speech recognition by converting the speech into syllable string first and then convert the syllable string into Chinese words. In the first stage of our recognizer, considering the small number of syllables in Chinese language, a frame synchronous stack decoder is used to integrate the syllable language model, so as to generate N-best list syllable strings. Also, fast matching technique is applied to reduce the time for recognition. In the second stage, considering the special monosyllabic wording structure in Chinese language, a syllable synchronous stack decoder is integrated with a word search tree and the word language model for the decoding of the best word sequence. In order to improve the recognition accuracy, an error correction model is also included to recover the syllable error decoded in the first stage. This thesis also proposes the above second stage is suitable for Chinese pinyin input. Specifically, the above second stage is a syllable to word decoder where a Chinese pinyin sentence can be converted into Chinese word sequence. Moreover, mixed language and non-Chinese syllable input is allowed either by treating the non-Chinese word as unknown word in the word search tree, or by putting the non-Chinese word in the word search tree for decoding.