Abstract: In query-by-humming systems，uncertainty can arise in query formulation due to user-dependent variability and machine-based errors.The paper proposes a new system，which using vowel detection for note segmentation.After that，the segmented humming audio is transcribed to symbols with pitch and duration information.At last，the transcribed audio is compared with the database and finds the closest melodic fragments.Experiment shows the retrieval accuracy is higher than 70% of best candidate with faster retrieval speed than traditional system，which achieves the requirement of practice application.

Key words: audio retrieval, music retrieval, vowel detection, feature representation with symbol

摘要： 哼唱的随意性和音乐特征提取算法误差都会影响基于哼唱的音乐检索系统的性能。针对上述问题，利用元音帧检测获得较为精确的音符边界，实现音符分割；对分割后的音符提取相对音高和音长，实现符号描述；最后将哼唱片段中音高和音长最值点周围的符号描述作为特征与数据库中的数据进行匹配，得到最相似的候选音乐。实验表明该方法对未经训练的哼唱者的首位匹配正确率达到70%以上，匹配速度也大大优于传统方法，检索性能基本达到了实际应用的需求。

关键词: 音频检索, 音乐检索, 元音检测, 符号特征描述