FEATURE-BASED EXTRACTION OF PLUCKING AND EXPRESSION STYLES OF THE ELECTRIC BASS GUITAR
Abeßer, Jakob
Lukashevich, Hanna
Schuller, Gerald
https://pdfs.semanticscholar.org/9818/3c01a5fbc7cba3598951509c2f0eac79d9a1.pdf
这篇是区分电贝斯拨弦技巧和expression styles的，用的方法非常信号处理。
先估每个谐波partial的能量包络，将partial分成attack和decay部分。随后在attack, decay或是真个音符层次上提了许多特征。
分类的方法使用了feature selection和feature space transformation - GDAf。经过space tranformation的feature，即使使用简单的分类器比如GMM和KNN也能有好的效果。
Very signal processing oriented work
They explore features, feature selection techniques and classifiers for the bass guitar plucking style and expressive style classification.
They first calculate PSD and estimate the envelope of each partial. The partials are divided into attack and decay parts.
Plucking style features and expressive styles features are extracted from either attack or decay or the entire note.
For the feature selection and feature space transformation, they use Inertia Ratio Maximization using Feature Space Projection (IRMFSP).
THE POTENTIAL FOR AUTOMATIC ASSESSMENT OF TRUMPET TONE QUALITY
Knight, Trevor
Upham, Finn
Fujinaga, Ichiro
http://ismir2011.ismir.net/papers/PS4-17.pdf
这个工作是评价小号演奏音符音色的。
收集了3个小号演奏者和1个长号演奏者的239个音符。问了7位评委用7尺度来评价这些音符，结果118个音符的intersubject相关性高，另外121个音符的相关性差，表示评委对其意见不一致。
提了56个特征，大部分都是频谱特征，用了SVM分类器。
文章有意思的地方是实验的设计。设计了：
（1）2, 3, 7个尺度的三个实验。
（2）leave one player out，训练集包括3个演奏者的音符，预测剩下的一个演奏者的音符。
（3）演奏者识别。
实验（2）的结果比（1）要差。由于模型可能fit是演奏者的特点而不是音色的好坏。比如演奏者4演奏的音符基本偏差，演奏者2的音符基本偏好，由其训练出的模型就可能会把音符同演奏者4想象的作为差的，同演奏者2想象的作为好的。
This work collected 3 trumpet players and 1 trombone player performance. Three levels of dynamics and 3 musical phrases are recorded. In total 239 notes.
They asked 7 raters to rate these notes, and take the round average score as the ground truth. The intersubject correlations are quite high for 118 notes but quite low for another 121 ones. The rating scale is from 1 to 7.
They used 56 features, most of them are spectral features. They used SVM as the classifier.
The best part of the paper is the experiment design.
(1) normal 2, 3, 7 classes classification. All players notes are mixed in the cross-validation folds.
(2) Leave one player out. Train on 3 players and test on another player.
(4) Identify the player experiment.
Results:
(1) 2 classes > 3 > 7, this is obvious
(2) leave one player out experiment is worse than (1). This is also understandable. Because the model might fit on the player's characteristics other than the tone quality. E.g. player 4 has more bad tones than good tones. Player 2 has more good tones than bad tones. The training set containing players 2 and 4 might be fitted by their player qualities. Thus, in the test set, if the player characteristics are more close to player 2, it will be classified as good. However, if it is close to player 4, it is bad.
HIERARCHICAL APPROACH TO DETECT COMMON MISTAKES OF BEGINNER FLUTE PLAYERS
Han, Yoonchang
Lee, Kyogu
https://pdfs.semanticscholar.org/ff70/498c98f7c7bf3a896acbc8781ea253d5aa26.pdf
这篇是好文。讲区分初级笛子演奏者的演奏错误的。
区分（1）笛子组装错误（2）fluctuated sound（3）错误的指位。
对每种错误都寻找合适的特征。比如组装错误会影响tuning，所以用HPCP peak histogram来找tuning center。
fluctuated sound导致inharmonic partials，所以用masked spectrogram来高亮这些partials.
所有找到的特征效果都很好。
Detect flute (1) assembling error (2) fluctuated sound (3) mis-figuring performance mistakes.
Assembeling error affects the pitch tuning, they use HPCP peak histogram to detect the tuning center.
fluctuated sound causes inharmonic partials. They use binary masked spectrogram to highlight these particals, then take the sum across the bins as the feature
Mis-fingering is represented by the timbre problems. So they used MFCC + sparse filtering + RF classication method.
All these features are quite effective for detecting corresponding mistakes.