Automatic Accompaniment Music Generation and Singing Skill Evaluation for Vocal Melodies

Abstract

Abstract Music composition may be one of the most di cult tasks for non-musicians. To allow nonmusicians to get a taste of music creation, we propose a system for generating accompaniment music for vocal melodies. In the system that we propose, a pitch detection module rst extracts the pitch classes from the vocal melody using a time domain auto-correlation method. According to these pitch classes, a singing skill evaluation module classi es the skill of the vocal melody as good or poor based on the pitch and the tempo. The pitch based singing skill classi er uses two features of pitch interval accuracy and classi es the singing skill using a trained Support Vector Machine. The tempo based singing skill classi er uses a vibrato suppression based onset detection technique to estimate the tempo of the vocal melody and classi es the singing skill. A Hidden Markov Model with learned probabilities then construct the best acceptable chord progression for the vocal melody by applying the Viterbi algorithm. The system then generates the accompaniment music using the constructed chord progression and combines it with the vocal melody to obtain the nal outcome. Finally, we present the results from the rst and the second study demonstrating that the performance of singing skill classi ers is fairly good with an accuracy of 81:8% and 87:5% respectively. We present results from the third study showing that our system is able to construct acceptable chord progressions for vocal melodies at a rate of 89:381% accuracy. The results from the nal study shows that the users are enthusiastic about the nal outcome of our system with above 3:76 mean score for all the Likert-scale questions.