PDF《CADENCE DETECTION IN WESTERN》

5 and 6). To investigate which musical parameters are of impor- tance for cadence detection, we perform an ablation study in which we subsequently remove certain types of features in order to evaluate the importance of the various kinds of features (Section 7). 2. DATA We perform all our experiments on the folk song collec- tion from the Meertens Tune Collections (MTC-FS, ver- sion 1.0), which is a set of 4,120 symbolically encoded Dutch folk songs.

1 Roughly half of it consists of tran- scriptions from ?eld recordings that were made in the Nether- lands during the 20th century. The other half is taken from song books that contain repertoire that is directly related to the recordings. Thus, we have a coherent collection of songs that re?ects Dutch everyday song culture in the early 20th century. Virtually all of these songs have a stanzaic structure. Each stanza repeats the melody, and each stanza

1 Available from: http://www.liederenbank.nl/mtc. consists of a number of phrases. Both in the transcrip- tions and in the song books, phrase endings are indicated. Figure

1 shows a typical song from the collection. The language of the songs is standard Dutch with occasionally some dialect words or nonsense syllables. All songs were digitally encoded by hand at the Meertens Institute (Ams- terdam) and are available in Humdrum **kern format. The phrase endings were encoded as well and are available for computational analysis and modeling. 3. OUR APPROACH Our general approach is to isolate trigrams from the melo- dies and to label those as either cadential or non-cadential. A cadential trigram is the last trigram in a phrase. We com- pare two kinds of trigrams: trigrams of successive notes (note-trigrams), and trigrams of successive pitches (pitch- trigrams), considering repeated pitches as one event. In the case of pitch-trigrams, a cadence pattern always consists of the three last unique pitches of the phrase. There are two reasons for including pitch-trigrams. First, pitch repetition is often caused by the need to place the right number of syllables to the melody. It occurs that a quarter note in one stanza corresponds to two eighth notes in another stanza because there is an extra syllable at that spot in the song text. Second, in models of closure in melody [11,15] suc- cessions of pitches are of primary importance. Figure

1 depicts all pitch-trigrams in the presented mel- ody. The trigram that ends on the ?nal note of a phrase is a cadential trigram. These are indicated in bold. Some cadential trigrams cross a phrase boundary when the next phrase starts with the same pitch. From each trigram we extract a number of feature val- ues that re?ect both melodic and textual properties. We then perform a classi?cation experiment using a Random Forest Classi?er [2]. This approach can be regarded a '

bag- of-trigrams'

approach, where each prediction is done inde- pendently of the others, i.e. all sequential information is lost. Therefore, as a next step we take the labels of the direct neighboring trigrams into account as well. The ?- nal classi?cation is then based on a majority vote of the predicted labels of adjacent trigrams. These steps will be explained in detail in the next sections. Figure 1. Examples of pitch-trigrams. The cadential tri- grams are indicated in bold. 4. FEATURES We represent each trigram as a vector of feature values. We measure several basic properties of the individual pitches and of the pattern as a whole. The code to automatically extract the feature values was written in Python, using the music21 toolbox [5]. The features are divided into groups that are related to distinct properties of the songs. Some features occur in more than one group. The following overview shows all features and in parentheses the value for the ?rst trigram in Figure 1. Detailed explanations are provided in sections 4.1 and 4.2. Pitch Features Scale degree Scale degrees of the ?rst, second, and third item (5, 1, 3). Range Difference between highest and lowest pitch (4). Has contrast third Whether there are both even and odd scale degrees in the trigram (False). Contour Features Contains leap Whether there is a leap in the trigram (True). Is ascending Whether the ?rst and second intervals, and both are ascend- ing (False, True, False). Is descending Whether the ?rst and second intervals, and both are de- scending (True, False, False). Large-small Whether the ?rst interval is large and the second is small (True). Registral change Whether there is a change in direction between the ?rst and the second interval (True). Rhythmic Features Beat strength The metric weights of the ?rst, second and third item (0.25, 1.0, 0.25). Min beat strength The smallest metric weight (0.25). Next is rest Whether a rest follows the ?rst, second and third item (False, False, False). Short-long Whether the second item is longer than the ?rst, and........