Frank Scherbaum

This paper describes the current status of a long-term project aimed at understanding the chordal syntax of traditional Georgian vocal music by analyzing sheet music in Western 5-line staff notation. As an important milestone, we present a generative grammar model based on the self-learning Kohonen model (Kohonen, 1989) in a prefix tree (Antonov, 2018; 2023) framework. This represents a significant improvement over the classical Markov model, as it allows for the influence of different context lengths for each chord in a chord sequence. We used this model to generate a large number of chord sequences, all conforming to the same grammatical production rules as our corpus. These were then used as training data for an artificial neural network to test whether, as in large language models (LLMs), ‘linguistic relationships’ could be identified by visually analyzing the embedding space of the network. The results for chord-to-chord relationships are inconclusive, as the spatial structure of the embedding map for individual chords cannot be interpreted unambigously. The embedding map for whole songs, however, shows a pronounced spatial clustering which reflects the different classes of our corpus. This suggests that the structure of the embedding map reflects the similarities and dissimilarities of the chordal syntax of the individual songs, which the network has learned in an unsupervised way.

This paper describes a feasibility study for the computational classification of traditional three-voiced Georgian vocal music, based on characteristic chord sequences extracted from digital scores. We demonstrate that for this purpose the differences between Western 5-line staff notation and a more appropriate heptatonic system for traditional Georgian music can be adjusted for by a simple transformation. A corpus of about 500 digital scores, consisting of labeled ‘song classes’, i.e. subsets of folk songs from different regions and of liturgical songs in different styles, served as a testbed for the development of a classification procedure based on a higher-order Markov model that - in addition to the classification - yields chord progression sequences for each ‘song class’. Their capacity for interpretation was tested by one hundred cross-validation runs, in which randomly selected subsets of ¾ of the size of each song class were used to train classifiers, which were then applied to the remaining ¼ subsets. The sizes of the intersections of the successfully classified songs in all cross-validations are interpreted as direct measures of the degree of representativeness of the songs for their respective ‘song classes’. Based on a second validation experiment, in which we split up the datasets into equally sized subsets of ½ and ¼ of the original subsets, respectively, we estimate that the smallest subset size for an interpretation of the observed chord progression patterns as properties of a ‘song class’ is of the order of 50 songs. Currently, in our corpus, this requirement is only met by the subsets from Svaneti and Shemokmedi.

Larynx microphones (LMs) make it possible to obtain practically crosstalk-free recordings of the human voice by picking up vibrations directly from the throat. This can be useful in a multitude of music information retrieval scenarios related to singing, e.g., the analysis of individual voices recorded in environments with lots of interfering noise. However, LMs have a limited frequency range and barely capture the effects of the vocal tract, which makes the recorded signal unsuitable for downstream tasks that require high-quality recordings. In this paper, we introduce the task of reconstructing a natural sounding, high-quality singing voice recording from an LM signal. With an explicit focus on the singing voice, the problem lies at the intersection of speech enhancement and singing voice synthesis with the additional requirement of faithful reproduction of expressive parameters like intonation. In this context, we make three main contributions. First, we publish a dataset with over 4 hours of popular music we recorded with four amateur singers accompanied by a guitar, where both LM and clean close-up microphone signals are available. Second, we propose a data-driven baseline approach for singing voice reconstruction from LM signals using differentiable signal processing, inspired by a source-filter model that emulates the missing vocal tract effects. Third, we evaluate the baseline with a listening test and further show that it can improve the accuracy of lyrics transcription as an exemplary downstream task.

This paper is concerned with how singers of Georgian traditional vocal music interact when singing together. Applying a variety of computational methods from audio signal processing and music information retrieval (MIR), we examine three existing corpora of (field) recordings for manifestations of a high degree of mutual coordination of the singers’ voices. We find numerous examples of harmonically controlled mutual intonation adjustments on both short and long time scales. Furthermore, we believe that the observed differences in melodic and harmonic scales can also be interpreted as (side) effects of the singers’ interaction with the possible goal of achieving harmonic togetherness (or consonance) on the time scale of individual (important) notes. In addition, together with the ensemble Khelkhvavi from Ozurgeti, we conducted an experiment demonstrating the synchronization of singers’ heartbeat rates during the performance of the Gurian song Chven Mshvidoba. The results of our analysis show that a variety of measurable signs of interaction between singers can be observed and documented in existing corpora of Georgian traditional vocal music. Our experience also shows that relevant information about the synchronization of the bodily functions of singers during performances can nowadays be obtained with reasonable technical and logistical effort even in a ‘real world’ framework, thereby allowing us to address questions related to the ‘ecological validity’ of this kind of measurement.

In this study we examine the tonal organization of the 2016 GVM dataset, a newly-created corpus of high-quality multimedia field recordings of traditional Georgian singing with a focus on Svaneti. For this purpose, we developed a new processing pipeline for the computational analysis of non-western polyphonic music which was subsequently applied to the complete 2016 GVM dataset. To evaluate under what conditions a single tuning system is representative of current Svan performance practice, we examined the stability of the obtained tuning systems from an ensemble-, a song-, and a corpus-related perspective. Furthermore, we compared the resulting Svan tuning systems with the tuning systems obtained for the Erkomaishvili dataset (Rosenzweig et al., 2020) in the study by Scherbaum et al. (2020). In comparison to a 12-TET (12-tone-equal- temperament) system, the Erkomaishvili and the Svan tuning systems are surprisingly similar. Both systems show a strong presence of pure fourths (500 cents) and fifths (700 cents), and ‘neutral’ thirds (peaking around 350 cents) as well as ‘neutral’ sixths. In addition, the sizes of the melodic and the harmonic seconds in both tuning systems differ systematically from each other, with the size of the harmonic second being systematically larger than the melodic one.

Three-voiced funeral songs from Svaneti in North-West Georgia (also referred to as Zär) are believed to represent one of Georgia’s oldest preserved forms of collective music-making. Throughout a Zär performance, the singers often jointly and intentionally drift upwards in pitch. Furthermore, the singers tend to use pitch slides at the beginning and end of sung notes. Musicological studies on tonal analysis or transcription have to account for such musical peculiarities, e.g., by compensating for pitch drifts or identifying stable note events (located between pitch slides). These tasks typically require labor-intensive annotation processes with manual corrections executed by experts with domain knowledge. For instance, in the context of a previous musicological study on pitch inventories (or pitch-class histograms) of Zär performances, ethnomusicologists tediously annotated fundamental frequency (F0) trajectories, stable note events, and pitch drifts for a set of 11 multitrack field recordings. In this article, we study how musicological studies on field recordings can benefit from interactive computational tools that support such annotation processes. As one contribution of this article, we compile a dataset from the previously annotated audio material, which we release under an open-source license for research purposes. As a second contribution, we introduce two computational tools for removing pitch slides and compensating pitch drifts in performances. Our tools were developed in close collaboration with ethnomusicologists and allow for incorporating domain knowledge (e.g., on singing styles or musically relevant harmonic intervals) in the different processing steps. In a case study using our Zär dataset, we evaluate our tools by reproducing the pitch inventories from the original musicological study and subsequently discuss the potential of computer-assisted approaches for interdisciplinary research.

This article is a companion paper to Scherbaum and Mzhavanadze (2020), Mzhavanadze and Scherbaum (2020) and Scherbaum and Mzhavanadze (2021). Together they describe the results of an interdisciplinary study of the three-part funeral dirge of Svan zär (‘zari’ in Georgian). In the present paper, we provide all the contextual (ethnographic, ethnological, historical) data that we collected from various written and/or oral sources and then processed and organized in a comprehensible way to help the reader better understand this unique phenomenon. We believe that in order to help answer the basic question of ‘how zärs work,’ the results of the acoustical, musicological, and phonetic (as well as language-music interface) analyses in the other three articles must be interpreted through an understanding of the full context of this unique cultural behavior.Thus, not only did we collect and integrate all the related data and review the literature, but we also raise and discuss some problematic issues, such as, for example, the etymology of the term zär. We also offer an interpretation/description of some conflicting historical data to generate interest for its further research. In addition, we introduce the reader to some details of our field research (2015, 2016), as they may be of great importance when discussing or interpreting the results of the analyses presented in other (analytical) parts of this interdisciplinary study.

Volcanic tremor signals are usually observed before or during volcanic eruptions and must be monitored to evaluate the volcanic activity. A challenge in studying seismic signals of volcanic origin is the coexistence of transient signal swarms and long‐lasting volcanic tremor signals. Separating transient events from volcanic tremors can, therefore, contribute to improving upon our understanding of the underlying physical processes. Exploiting the idea of harmonic–percussive separation in musical signal processing, we develop a method to extract the harmonic volcanic tremor signals and to detect transient events from seismic recordings. Based on the similarity properties of spectrogram frames in the time–frequency domain, we decompose the signal into two separate spectrograms representing repeating (harmonic) and nonrepeating (transient) patterns, which correspond to volcanic tremor signals and earthquake signals, respectively. We reconstruct the harmonic tremor signal in the time domain from the complex spectrogram of the repeating pattern by only considering the phase components for the frequency range in which the tremor amplitude spectrum is significantly contributing to the energy of the signal. The reconstructed signal is, therefore, clean tremor signal without transient events. Furthermore, we derive a characteristic function suitable for the detection of transient events (e.g., earthquakes) by integrating amplitudes of the nonrepeating spectrogram over frequency at each time frame. Considering transient events like earthquakes, 78% of the events are detected for signal‐to‐noise ratio = 0.1 in our semisynthetic tests. In addition, we compared the number of detected earthquakes using our method for one month of continuous data recorded during the Holuhraun 2014–2015 eruption in Iceland with the bulletin presented in Ágústsdóttir et al. (2019). Our single station event detection algorithm identified 84% of the bulletin events. Moreover, we detected a total of 12,619 events, which is more than twice the number of the bulletin events.

This paper reports on the results of a follow-up study to Scherbaum and Mzhavanadze (2020) and Mzhavanadze and Scherbaum (2020), which jointly describe the acoustical and musicological properties, respectively, of a new collection of field recordings of three-voiced Svan funeral dirges, known as zär in Svan and zari in Georgian. The focus of the present work is on language–music relation and phonetic properties. It was motivated by the pioneering work of Bolle Zemp (1994; 1997; 2001), who for the first time examined this topic from both an ethnomusicological and a linguistic perspective. We revisit some of Bolle Zemp’s observations and assumptions with a substantially expanded (by a factor of 11) data set and by using computational phonetic analysis tools not available at the time of her study. Her observation of correlations between pitch, duration, and timbre features related to the interjection of woj (wai) is consistent with our analysis for only some of the singers. Therefore, rather than assigning a general semantic meaning to these correlations, we offer an alternative interpretation as a natural consequence of the formant tuning techniques, which we found employed in the vocal production of some of the singers.

This paper is a companion paper to Mzhavanadze & Scherbaum (2020). Jointly, the two papers describe the results of an interdisciplinary study of three-voiced Svan funeral dirges, known as zär in Svan and zari in Georgian. In the present paper, to which we refer as paper 1, we analyze the musical acoustical properties of a new set of field recordings collected during an ethnomusicological field expedition to Georgia in 2016. The aim of the study is to investigate the tonal organization of eleven different performances of six different variants of zär, performed by singers from different villages. For some of the performances, we observe a strong gradual pitch rise of up to 100 cents per minute. The intra-variant differences in the performances of different groups of singers were observed to be remarkably different, including the use of significantly different harmonic tuning systems. In contrast, two subsequent performances of the Mest’ia variant of zär by a group of singers recorded in Zargǟsh were essentially identical. This demonstrates the widespread absence of improvisational elements in these two performances. One of the most interesting results of our analysis is the observation that the musical structure of zär, expressed, for example, in its ambitus, the complexity of its melodic progression, and its harmonic chord inventory, change systematically along the course of the Enguri valley.