Hmm's Humming

Computer search audio clips can be based on the title or file name, but the description of file name or the title is not integrity and subjectivity, it is difficult to find audio clip to meet the specific requirements. Query by Humming ,as a search of the most natural way to retrieve the music has been concerned about. HMM as a statistical model, which has solid theory, has been introduced in the field of music search. However, most of the studies built model based on note , note as a unit to deal with training and recognition, through the segmentation to obtain the notes of each melody tone of the sound, which lead the humming way is unnatural.To deal with the problem, the topic studied in-depth on the existing research results about the humming theory and the HMM model, analyzed the feasibility and superiority theory of HMM model in query by humming, improved the model and related algorithms, designed the prototype system to retrieve the music and so on. By melody signal for a short time frame, extracted the feature vector of melody, studied the characteristics of these vector sequence in the statistics. before retrieve ,each melody must be trained and modeled.and then calculate the probability of test melody based on the models. select the maximal probability as a result of retrieve. realized an effective music retrieve.Specific activities include the following: 1. analyzed the structure of the music retrieve,came out the express methods of the characteristics of the audio information and MIDI music tune.2. analyzed the basic principles of HMM model,outlined the three core questions in the model,came out the three basic algorithms about these questions,put forward the theoretical feasibility and advantages in humming feature extraction and recognition.3. analyzed the process of training and identifying certain types of objects using the HTK toolkit,used HMM as a classify-ware, humming can be classified, settled the classification algorithm based on LFPC.4. discussed the selection and extraction of characteristics of MIDI, with sequence Hidden Markov Model (CHMM) foundation eatablish model of melody, frame units for training and identify, through experiment verified the effective of the HMM model for humming search.