Identification of Songs using Humming Quer

Abstract

Research focused on ‘pure approach’ such as ‘sound engineering’ to query humming part of sound against music or song to recognize similarities using distance variance. With the increase number of music files on different types of devices it’s very difficult to locate a music file in instant. Thus, audio file content is not visually understandable by humans without any expert help. In this study it has been conducted research on how to the reading of this human understandable streams of music to match with respective hummed input. As for the proposed method in this study it uses query by humming model that can predict song for user humming. In this model it has conducted in many three areas. Which is convert humming wave to frequency vector using ACF based approach. Then further quantify these frequency vectors using hop windows for better frequency wave based on beat of the song. This frequency vectors related average filtering is also discussed in the study where generalized frequency vectors into more informative beat frames-based frequency vectors. Secondly conversion of Frequency vector to Note vector has been conducted. These notes have been gone through several filtering mechanisms and has used note difference calculation to provides note difference vectors which will be used in third step where measure DTW difference calculation to match Hummed queries and real songs. Further study has conducted research on Quantified Dynamic Time Wrapping (QDTW) approach to carry out local minimum distance analysis in manual approach. By conducting research on these areas, it has proven that sound engineering approach can give accurate results when identifying songs for hummed queries over 83% overall accuracy for relatively better pitched Hummings respect to 67% accuracy for selected bad set of Humming data. Since research carried out mainly on identifying Sinhala songs humming data used were created by school of professional singers at perfect pitch, and random pitch, which has been evaluated separately to distinguish performance of the model over giving results ranking at top 5 results scale.