Computational Analysis of Yaredawi YeZema Silt in Ethiopian Orthodox Tewahedo Church Chants

The Ethiopian Orthodox Tewahedo Church (EOTC) chants hold immense cultural and religious significance in Ethiopia, yet they are largely overlooked [1].¹¹1The Eritrean Orthodox Tewahedo Church, which separated from the EOTC administration system a few decades ago, also utilizes these chants. We acknowledge its important role in preserving this sacred form of church music. The EOTC chant is believed to have originated with Saint Yared (505–571), who composed the three EOTC chanting modes (YeZema Siltoch in Amharic language),²²2Siltoch is the plural form of silt. For simplicity, the Amharic phrase YeZema Silt and the English phrase chanting mode will be used interchangeably throughout this paper. namely Ge’ez,³³3The term Ge’ez holds various connotations depending on context; here, it denotes one of the three chanting styles. Conversely, it also refers to the language and may have other applications. Ezil and Araray. Saint Yared’s pioneering musical compositions, liturgical chants, and associated dance movements had a significant impact on Ethiopian sacred music tradition [2]. The Debterawoch (also called Merigetawoch), who are the expert musicians and heirs of Saint Yared, play a crucial role in the transmission and performance of the sacred music [1]. Ethiopian sacred music has been preserved through oral and written traditions, with written documents supporting and reinforcing the ongoing oral traditions [3]. The significance of the EOTC chants in Ethiopian culture and worldwide is evident through the two major spiritual mass celebrations that have been recognized by UNESCO as intangible cultural heritages: the Commemoration Feast of the Finding of the True Holy Cross of Christ (in 2013) and the Ethiopian Epiphany (in 2019).⁴⁴4https://ich.unesco.org/en/state/ethiopia-ET?info=elements-on-the-lists These two celebrations, primarily accompanied by the EOTC chants, are among the five intangible cultural heritages from Ethiopia registered by UNESCO.

Despite its long history and development, the research of EOTC chants was quite rare. Among them, a renowned ethnomusicological work from Western academia was by Shelemay et al. [3, 4], based on the analysis of a series of EOTC chants collected in Addis Ababa, 1975. They discussed the oral and written tradition of EOTC chants, the EOTC chant music notation system, and further the definition of the three chanting modes, specifically the pitch sets used in each of the modes. It should be noted that in this work, all the recordings were transcribed and analyzed by ear. As stated in the paper, the analysis, for a limited number of recordings, was sometimes challenging when transcribing the non-Western music scales. With no indigenous classification of their pitch materials [3], YeZema Siltoch remains a primary research topic in the music theory of EOTC chants.

This paper is a study on YeZema Siltoch of the EOTC chants from computational perspectives. Our contributions in this paper are three-fold. First, we propose a new dataset for YeZema Silt classification and analysis. Second, we benchmark the YeZema Silt classification on the dataset using neural network classifiers with a number of features, primarily the pitch contour features which have been verified useful in analyzing various kinds of music [5, 6, 7, 8, 9, 10, 11]. Third, we perform a comparative study with [3, 4] to echo, and to revise their statements as well: while the pitch sets used in Ezil and Araray was regarded as the same [3], our numerical results indicate notable difference in between them. In the rest of this paper, we will have a background introduction of EOTC chants in Section 2. The proposed dataset, benchmarks and the comparative study will be in Sections 3, 4, and 5, respectively. Conclusion and future works will be given in Section 6.

The dataset was manually collected from the Eat the Book website,⁷⁷7https://eathebook.org/, We acknowledge the website’s administrators for their invaluable contribution. a hub of numerous audio books for most of the teachings in the EOTC school departments, with full and partial coverage. From the available audio books, we selected the Se’atat Zema (Horologium chant), which is part of Qidase-bet department. All the audios selected for our dataset were recorded by a single scholar at a sampling rate of 44,100 Hz and in stereo channel.

Our first step in the audio arrangement process involved narrowing the gap between the longest and shortest duration among the audios. Long audios, such as those over 13 minutes, were segmented into shorter audios of less than three minutes (180 seconds) in a way that preserves meaningful segments. This segmentation process also applied to audios that contained multiple chanting modes. For example, if an audio had 160 seconds of Araray mode followed by 22 seconds of Ezil mode content, it would be segmented into two separate audios of 160 seconds and 22 seconds. Recordings that were less than three minutes but still had multiple modes were also segmented based on the respective duration of the included chanting modes.

On the other hand, short audios, like a 19-second audio, were merged with neighboring context audios when applicable to our assumptions. If no neighboring audio with the same mode was found, it would be counted as a separate audio. As we arranged all audios to be in a single mode, we have a corresponding mode label for each audio. Another audio cleaning process was removing non-chant segments as the recordings included short explanatory statements about the corresponding chants. We manually removed them to ensure that the full audio content will be for chanting. In this process we also have shortened the duration of significant silent regions, resulted in only two silent regions above two seconds, particularly 2.25 and 2.14 seconds. After such cleaning procedures, the overall duration distribution of our dataset, ranging from 20.142 seconds to 177.476 seconds, is shown in Figure 2. As our immediate future work, we are working on expanding our dataset by including annotation of word-level lyrics to audio alignment as well as other features, which are not uncovered in this paper to keep the focus. We will do more research regarding other possible additional features.

Table 1 presents the comparison between the previously used dataset (i.e., the recordings collected by Shelemay et al. in [4]) and our dataset. While the previous dataset, with a total of 24 instances, is less than one hour, our dataset accounts for more than 10 hours, with a total of 369 instances. The chanting mode annotations of the dataset are available on https://github.com/mequanent/ChantingModeClassification.git.

As a preliminary study, we only consider using time-averaged audio features (i.e., the features ignoring the information lying in the temporal dimension) for Yaredawi YeZema Silt classification. Focusing on such features also supports our subsequent discussion on the pitch distributions of different chanting modes (see Section 5).

The goal of our analysis of YeZema Silt is using computational tools to individually identify the pitches utilized in the three chanting modes. Any attempt to this relies on some music theoretical assumptions. The classification results presented in Section 4.3 supports two assumptions that facilitate the analysis: first, YeZema Silt is a song-level property that can be satisfactorily described with time-average pitch distributions; second, YeZema Silt can be identified by a classifier invariant to pitch-shifting (i.e. convolution). On the other hand, the classification results also expose a few technical limitations. While the raw pitch distribution (i.e., without pitch contour stabilization) yields the best classification accuracy, it is highly noisy and therefore less applicable for our analysis purpose. In fact, we found in our study that the raw and the morphetic pitch distribution are relatively deficient in the below-mentioned analysis process. Therefore, instead of advocating a specific setting in terms of classification accuracy, we decided to use the calibrated pitch contour with masking stabilization method on the full length audio for subsequent analysis, although its performance is not the most favorable. It is worth noting that in this case, the performance gap between within-dataset CV and cross-dataset is relatively small among all settings.

Our approach, which partly resembles [10], contains three steps: 1) shift the pitch distributions of each recording such that each of them are best correlated (i.e., best aligned); 2) compute the average of the aligned pitch distribution for all the recording of the same chanting mode; 3) employ the Gaussian Mixture Model (GMM) to estimate the representative pitch set from the distribution.

The top row of Fig. 3 illustrates the aligned pitch distributions for the three chanting modes and the two datasets. We observe that the recordings in the same chanting modes typically have similar pitch distributions over the two datasets. Such a consistent trend is also observed from the average pitch distributions (middle row of Fig. 3), which shows that only one pitch (the third peak from the left) from the two dataset in Araray is somehow different.

The bottom row of Fig. 3 shows the GMM-estimated pitch distributions for all the recordings from both datasets. By selecting the pitches summing up to maximal weights within one octave, we obtain three representative pitches for Ge’ez (denoted as $g_{1}$, $g_{2}$ and $g_{3}$, from low to high), five for Ezil (denoted as $e_{1}$, $e_{2}$, $e_{3}$, $e_{4}$ and $e_{5}$) and also five for Araray (denoted as $a_{1}$, $a_{2}$, $a_{3}$, $a_{4}$ and $a_{5}$).⁸⁸8Here, the subscript number does not imply the hierarchical order of the musical scale (e.g., $g_{1}$ does not mean “the tonic of the Ge’ez mode”). The hierarchy of these pitches is another research question and will be considered as future work. Other representative pitches outside this octave are also notated: the pitch being one octave below $g_{3}$ is denoted as $G_{3}$, while the pitch one octave above $g_{1}$ is denoted as $\dot{g_{1}}$. The same naming rules also apply for Ezil and Araray.

Table 4 shows the GMM-estimated parameters for the three modes. First, the pitches used in the Ge’ez mode are more flexible than other two modes, as can be observed by their variances than the pitches in other two modes. Among them, only $g_{2}$ has the variance less than 10 cents. $g_{3}$ and $g_{2}$ form a major third ($\Delta\mu=400$ cents) while $g_{1}$ and $g_{2}$ form approximately a minor third $g_{2}$ ($\Delta\mu=324$ cents). The pitches of $g_{1}$ and $g_{3}$ can vary by more or less semitones. Besides, we also observe that the octaves of $g_{1}$ and $g_{3}$ (i.e., $G_{3}$ and $\dot{g_{1}}$) also have large variances. This implies that such variance (flexibility of pitch) depend on the pitch name rather than the register. These findings are basically in line with the statements (a scale $\sharp g$-$a$-$\flat b$-$c^{\prime}$-$\sharp d^{\prime}$-$e^{\prime}$-$f^{\prime}$ while $g$-$c^{\prime}$-$e^{\prime}$ are the stem pitches) made in [4].

Both the Ezil and Araray modes have five representative pitches within one octave. However, the five representative pitches of them are different. For Ezil, all the intervals lie between 200 cents (major second) and 300 cents (minor third), while for Araray, the intervals distribute from 172 cents (less than a major second) to 347 cents (in between a minor third and a major third). In other words, there is a consistent trend that the intervals in Ezil are more equally distributed than Araray. There are also some flexible usage of pitch, for example, $e_{5}$ ($E_{5}$) in Ezil. These suggest that the pitch sets found in [4] needs revision: from our observation, each of the pitch sets used in the three EOTC chanting modes is distinctive. Besides, a mode is characterized by not only its pitch centers, but also its pitch variances.

In this paper, we presented a research on a relatively underexplored music genre, the Ethiopian Orthodox Tewahedo Church (EOTC) chant, from three computational perspectives. First, through a rigorous data cleaning and annotation process, we presented a new and high-quality EOTC chant dataset, which can be extended for various music information retrieval (MIR) and music generation tasks. Second, we conducted a chanting mode (YeZema Silt) recognition task using our dataset and achieved promising results. Additionally, this paper is, to our knowledge, the first to computationally analyze the pitch sets of the EOTC chanting modes, specifically YeZema Siltoch, with new musicological insights. In the future, we plan to keep enriching the annotations of the datasets, by incorporating more details like lyrics, chanting options, reading tones and other potential features. Analyzing YeZema Siltoch using the features in the temporal dimension and the new data annotations are also our ongoing projects.

The EOTC chants encompass a wide range of styles and forms. In this paper, we specifically concentrated on the Se’atat Zema (Horologium chant), which falls under the Qidase-bet department. Our objective is to encourage responsible research on EOTC chants, as computational research in this area can lead to technological advancements that enhance the learning process and increase accessibility. Diversifying the data and MIR of EOTC chants for the protection and promotion of this spiritual-cultural heritage is also our future work in the long term.

The chants we are working on belongs to the Ethiopian Orthodox Tewahedo Church (EOTC). The oral traditions and beliefs that mainly preserve these chants for more than 1500 years should be credited properly. This work aims to contribute on promoting the chants and finding solutions for their easy understanding. There is no intention to modify the chants in any form.

No explicit permission request is done to the Eathebook.org administrators as recordings are open to the public education and we are using for educational purpose. We value their great effort in making the chants publicly available.