Violin Fingering Estimation Based on Violin Pedagogical Fingering Model Constrained by Bowed Sequence Estimation from Audio Input

This work presents an automated violin fingering estimation method that facilitates a student violinist acquire the “sound” of his/her favorite recording artist created by the artist’s unique fingering. Our method realizes this by analyzing an audio recording played by the artist, and recuperating the most playable fingering that recreates the aural characteristics of the recording. Recovering the aural characteristics requires the bowed string estimation of an audio recording, and using the estimated result for optimal fingering decision. The former requires high accuracy and robustness against the use of different violins or brand of strings; and the latter needs to create a natural fingering for the violinist. We solve the first problem by detecting estimation errors using rule-based algorithms, and by adapting the estimator to the recording based on mean normalization. We solve the second problem by incorporating, in addition to generic stringed-instrument model used in existing studies, a fingering model that is based on pedagogical practices of violin playing, defined on a sequence of two or three notes. The accuracy of the bowed string estimator improved by 21 points in a realistic situation (38% → 59%) by incorporating error correction and mean normalization. Subjective evaluation of the optimal fingering decision algorithm by seven violinists on 22 musical excerpts showed that compared to the model used in existing studies, our proposed model was preferred over existing one (p = 0.01), but no significant preference towards proposed method defined on sequence of two notes versus three notes was observed (p = 0.05).