【作者】 袁斌； 许洁萍；

【Author】 YUAN Bin, XU Jie-Ping(Department of computer, School of information, Renming University, 100872, China)

【机构】 中国人民大学信息学院计算机系；

【摘要】 本文对利用HMM模型进行乐音哼唱检索系统进行了研究。与已有的哼唱检索系统不同,本文对用户哼唱输入的音高差及音长比进行了统计分析,合成产生了符合用户哼唱输入的旋律及节奏训练数据库;根据对midi数据库的分析结果,对HMM模型中特征值的选取进行了新的组合、减维,对HMM模型训练算法提出了改进,并通过实验验证了HMM模型进行音乐哼唱检索的有效性。在含有1500个音乐片段的数据库中,27个检索段的前5位命中率为85.2%,取得了可喜的结果。更多还原

【Abstract】 HMM is a kind of model designed to handle random phenomenon. Because of its solid mathematics theoretical foundation, it is widely used in pattern-recognition, biological information DNA index and speech recognition fields. In recent years, HMM is applied to the music information retrieval field too. In 90s, Jonah Shifrin evaluated the retrieval capability of HMM as to different synthetic query inputs. The result of this study showed that HMM method can better tolerance the missing note and humming error of rhythm.We studied the music humming retrieval system utilizing HMM models. It is different from the existing systems. At first, we analyzed a midi database which include more than1000 songs and 2 millions notes and got the distribution results of delta pitch and IOI (Inter Onset Interval) ratio. From these distribution rules, we found that delta pitch of notes could be quantified into 15 levels and IOI ratio of notes could be quantified into 12 dispersed spots. In the experiment, we chose delta pitch and IOI ratio as the state of HMM models. Second, based on the analysis results of the midi, we chose a new association, reduced linking to characteristic value in HMM models and proposed improving methods to the training algorithms of HMM models. Finally, we collected 27 phrases of 6 users, segmented and labeled manually in order to study the input error rules of users’ singing. Through a statistical method, we got some useful rules about the habit of users humming and built a training database according to the user’s humming melody and rhythm rules. The training database included 100 themes and 359 phrases. A theme was corresponding with a HMM models and a model used 10 to 15 training sequences. Through experiments in the databases containing about 1500 music sections, the top 5 of 27 retrievals are 85. 2%. This result is inspirer and proved that the validity of music humming system base on HMM models. In the future, we will study theinput model of users’singing and add note duration ratio with certain weight to the state of HMM models.更多还原