【摘要】 为快速、准确地判断语音流中是否含有关键词，提出了一种基于LSTM两步检索的关键词识别系统。将连续语音流分割成独立音节，采用过零率直方图进行初步检索，基于过零率直方图的相似度比较的计算量小，可快速排除非关键词。对初检时判断为关键词的音频片段进行精检，使用基于LSTM的分级系统进行音素识别，通过贪心搜索算法解码以确认是否为目标关键词。仿真结果表明，基于LSTM的网络能更有效地提取音素特征，基于两步检索LSTM的关键词识别系统计算量小、速度快、识别率较高，且易于动态扩展目标关键词，具有较好的实时性。更多还原

【Abstract】 A two-step keyword recognition system based on LSTM is proposed to quickly and accurately determine whether there are keywords in speech stream. Firstly, the continuous speech stream is segmented into independent syllables, and then the zero-crossing histogram is used for initial retrieval. The similarity calculation based on the zero-crossing histogram computes less, and non-keywords can be eliminated quickly. Then the speech fragment which is judged as the key word is carefully examined. The phoneme is recognized by the classification system based on LSTM, and then the speech fragment is confirmed as the keyword by greedy search algorithm. The simulation results show that the LSTM network is more effective in extracting phoneme feature. The two-step keyword recognition system based on LSTM has the advantages of small computation, high speed, high recognition rate, easy dynamic expansion of target keywords, and good real-time performance.更多还原