| Solfege refers to the solfeggio who converts the notes in the score into corresponding tones and sings when the solfeggio sees a music score.For beginners,it is usually required that the solfeggio notes be complete and pitch accurate.However,the traditional assessment method of Solfeggio Teaching relies heavily on manual work,and it is impossible to realize batch and automatic assessment.Most of the existing automatic assessment systems are subjective assessment of the overall output of music,but lacks the objective feedback of note-level correctness.Therefore,it is of great significance to study the objective and notelevel automatic assessment system of solfeggio.The existing automatic solfege assessment system usually inputs a piece of audio.The system extracts audio features such as pitch,volume and melody of the solfege audio,and compares the features with the standard audio,and finally outputs the solfege assessment.The automatic solfege assessment system designed and implemented in this paper inputs the sequence of the vocal audio and the standard musical notes.The system divides the solfege audio according to the notes to obtain the sequence of the vocal notes,and the sequence of the vocal notes and the sequence of the standard scores.Align the sequence of solfege notes with the standard music notes,and compares the pitches one by one,and finally outputs the correctness assessment of the note level.The system proposed in this paper involves three technical points: onset detection,pitch extraction and note sequence alignment.This paper focuses on the problem of human voice onset detection and note sequence alignment,and implements the corresponding algorithm combined with sample analysis.The main innovations of this paper are shown as follows:(1)In onset detection,aiming at the low accuracy of the onset detection method based on traditional acoustic model,training an onset detector using CNN,which can reach a F-score at 90.58% on HUST-Solfege,much superior to non-deep-learning methods;(2)In note alignment,aiming at the fact that the previous feature matching algorithms used absolute pitch,interval and onset of notes as features can not solve loss or add note,this paper creatively apply Needleman-Wunsch algorithm,an alignment algorithm famous in bioinformatics,to align between detected pitch and music score by modeling on relative relationship of note sequence;(3)In system and assessment model,this paper designs and implements an automatic solfege assessment system,and proposes a ternary assessment model to assess the Correct Rate of Solfege(CRoS)of solfege samples and the accuracy of solfege detection system,it also proposes and public self-built solfege dataset HUST-Solfege to everyone.The solfege assessment system designed and implemented in this paper can reach an average accuracy at 93.68% on HUST-Solfege,which achieves application-level.It can provide a comparison benchmark of the follow-up research of the automatic solfege assessment system. |
