Singing is a worldwide activity across all walks of life. Many of us have had Karaoke defeats of singing a beloved song but making the others plug their ears. A major reason for such failure is that the songs one loves might not fit his/her personal laryngeal anatomy which determines the capacity of one's voice source. myDJ is a karaoke recommendation system which recommends proper songs based on the pitch and intensity that one can nicely produce. Unlike most state-of-the-art recommendation system-s, which rely on the similarity defined for the contents of songs, the listening habits of users, or the listening patterns, myDJ is the first prototype which recommends songs according to one's physical phonation area. The key challenge in this approach is to connect the singer's phonation model, which is typically described as a Vocal Range Profile (VRP)[1] in clinical quantitative voice assessments, to the music database, so that a given profile retrieves suitable songs. Unfortunately, a VRP is not enough to retrieve songs, as it only describes the minimum and maximum sound pressure levels (dB) across the singer's vocal range (Hz) without evaluating the voice quality. Our work focuses on techniques to build such connection for retrieval. myDJ consists of four modules: (1) The singer profiler, which creates a profile for each singer via a tranditional musical test known as Messa di voce. The singer profile consists of two parts, the VR-P which is depicted as a 2D region, and the Overall Voice Quality (OVQ)[3]combining multiple voice quality features as a distribution over the 2D region defined by VRP. (2) The Midi Song Database. We consider each song in the database as a documen-t, and the notes in a song at the same pitch or intensity as a pitch term (IT) or an intensity term(IT) respectively. The duration of each IT and the TF-IDF weighted duration of each PT, which are accumulated by the note duration in a song, are the major factors affecting the fitness of a song. Thus, we define a song's profile as the combination of the TF-IDF weights, the PT and IT durations, and the pitch and intensity values of all notes in it. (3) The learning to rank module, where algorithm Listnet[2] is applied. (4) The song recommendation module, where songs are recommended using the ranking function learned by Listnet. myDj works in two phases: 1. The offline training phase: The learning algorithm uses (1) the singer profiles, and (2) the five-level fitness scored song profiles which are manually labeled by the singers for themselves, as the training data. During the training process, each singer profile is considered as a query and the song profiles as the documents. Thus, the features are extracted from all (query, document) pairs, and Gradient Descent is applied to calculate the parameters in the ranking function. 2. The online phase: First, the test subject has to take the musical test to acquire the singer profile. Next, feature extraction will be conducted for all (query, document) pairs. To expedite this process, we utilize a document index to prune documents (songs) before any feature extraction. Finally, the rank score of each song is calculated using the ranking function learnt in the first step. Figure 1 shows a screen shot of myDJ. The left part of the interface shows a singer profile, where the 2D region indicates the VRP and the color of each pixel indicates the OVQ value (Red for good and blue for poor). The top right screen shows a ranked list of songs being recommended for this subject.
展开▼
