Deep Composer Classification Using Symbolic Representation

We used MIDI-only archive of the MAESTRO dataset v2.0.0 [4]. It provides over 200 hours of classical MIDI performances along with metadata such as composer, title, and duration. We excluded pieces with multiple composers to remove any ambiguity and selected composers with more than 16 pieces. These steps left 505 pieces by 13 composers in our experiment. Data are split into 7:3 train and test sets (347 and 158 pieces respectively) using stratified sampling with respect to the composer label.

In the dataset, each midi recording corresponds to a piece. To deal with the variable durations, only a randomly selected segment of the whole recording is fed to the model. The duration of a segment is set to 20 second (400 bins). For both training and inference of the model, we use 90 segments that are uniformly sampled over time.¹¹1All the details of the experiment including model and dataset are released at https://github.com/KimSSung/Deep-Composer-Classification.

After a preliminary experiment, we chose ResNet-50 as a reference model based on its performance after testing ResNet-18/34/50/101. The model was trained using SGD with momentum and weight decay. We used CE loss function with cosine annealing for LR scheduling.

We report our reference model with ‘weighted’ averaged F1 score of $0.8333$.²²2ResNet-18/34/101 achieved F1 scores of $0.7962$/$0.7881$/$0.7892$. This can be loosely compared to [5, 6] although the problem definition and the dataset vary. [5] using gradient boosting and CNN achieved F1 scores of $0.742$ and $0.700$ for classifying 15 and 6 classical composers, respectively. Similarly, a ConvRNN-based system in [6] achieved 70% accuracy for classifying 6 composers.

Based on this reference model, we investigated the performance variation by composers. First, the Spearman rank correlation coefficient of the composers’ birth years and performances was $-0.45$, i.e., the model performed better for relatively old classical composers. This results seem musically plausible and related to the fundamental property of the problem – Because the style of music has been diversified and more complicated chords have been popularized over time, recognition of composer by the music may be inherently more difficult for relatively modern composers. Second, we also hypothesized if the label imbalance in the training data has an effect on the performance, but the Spearman rank correlation coefficient between the number of training items (pieces) and performance was $-0.13$, which does not seem significant.

The confusion matrix in Figure 3 shows that a majority of the misclassifications occurred within the same era (Denoted with bold boxes). Out of 19 misclassified pairs, only 5 were classified as composers from different era. We assume the composers from the same era are more likely to share similar musical patterns, and this indicates that the model successfully learned such patterns during training.

Number of segments: We trained various instances of the model by increasing the number of segments from 5 to 90 as reported in the first column of Table 2. The result indicates that segments over 30 provide information that is diverse enough for the model to perform at a certain level (over 80% of F1 score).

In this paper, we showed that a composer classification can be done using symbolic representation. We investigated the trained model by varying the within-piece coverage and input data representation as well as analyzing the performances per composer. A future direction may include separating the influence of players, generalizing towards other instruments, and domain adaptation to other genres.