Exploring the Classification of Pipe Organ Sounds: Recognizing Instruments, Organ Types, and Registrations

Exploring the Classification of Pipe Organ Sounds: Recognizing Instruments, Organ Types, and Registrations Vince During1,PabloDumenil 2,andPetervanKranenburg 1[000000029760242] 1Utrecht University, Utrecht, Netherlands, {v.during,p.vankranenburg}@uu.nl https://www.uu.nl/en/research/interaction/music-information-computing 2Utrecht, Netherlands, [email protected] Abstract. Classification of pipe organ sounds is challenging as the pipe organ has a huge variety of timbres depending on the selection of stops and the type of the instrument. Recent advances in foundational models for audio enable the exploration of classification tasks for pipe organ audio recordings. We focus on Organ Type Recognition, Organ Identification, and Registration Recognition (both intra-track and for full tracks). Employing BYOL-A embeddings and the multi-layer perceptron, we achieve impressive classification results. Keywords: Instrument Recognition ·Pipe Organs ·Sound Embedding. 1Introduction The pipe organ is one of the oldest and most complex musical instruments, with ahistoryspanningovertwomillennia.DuringtheRenaissanceandBaroque periods, the organ underwent significant improvements in design and complexity, with composers like Johann Sebastian Bach enhancing its prominence through their compositions. It became a symbol of religious and civic pride, commonly found in churches, cathedrals, and concert halls. In the nineteenth and twentieth centuries, its versatility and complexity grew, resulting in large symphonic organs designed to emulate the sound of full orchestras. The many timbres that can be realized on a pipe organ pose a challenge for automatic processing of organ sounds. Extensive work has been done on physical modeling of pipe organ sound [6,8,4]andonsynthesis[2,5,1], but to the best of our knowledge, common Music Information Retrieval tasks [9], including instrument recognition, source separation, automatic transcription, and timbre recognition have not yet been explored for the pipe organ. For these tasks, a rich model of sound is required that could potentially capture the many nuances in the timbres of the organ. Until recently, this was All rights remain with the authors under the Creative Commons Attribution 4.0 International License (CC BY 4.0). Proc. of the 17th Int. Symposium on Computer Music Multidisciplinary Research, London, United Kingdom, 2025 Proc. of the 17th International Symposium on CMMR, London, UK, Nov. 3-7, 2025 298 V. During et al. amajorimpediment,buttheintroductionofgeneralpurposeaudioframeworks provides a promising foundation. In a previous stage of our project [3], we found that the Bootstrap Your Own Latent for Audio (BYOL-A) framework [7]offersa rich audio representation that is suitable for the task of registration recognition. In this paper, we explore further the extent to which the BYOL-A framework can be used for the following four tasks: Organ Type Recognition. We classify organs into broader types (Renaissance, Baroque, Romantic and Symphonic) solely based on their sound. Organ Identification. We automatically identify individual organs from their sound. Each organ in the dataset is treated as a separate class. Full Track Registration Recognition. We identify the registration (specific combination of stops or timbre) used throughout an entire track, building on prior work [3]. Intra-Track Registration Recognition via Slicing. We detect registration changes within a single track by analyzing time-sliced audio segments, creating a timeline of registration transitions. Solutions for these tasks would be extremely helpful for managing large audio collections or for playlist generation. This work was done in collaboration with OrganRoxx3,anonlinestreamingservicefororganmusic.OrganRoxxhostsone of the largest collections of recorded organ music in the world, including digitized audio as well as a wealth of meta data. For a streaming service, organizing the collection and meta data is a crucial requirement to be able to construct engaging and balanced play lists. Currently, for many tracks, important information is lacking. This includes the timbre of the music, as well as the type of organ on which the music is performed. In this paper, we address both aspects. 2BackgroundonRegistrations Registration in the context of the pipe organ refers to the selection and combination of different sets of pipes to produce various timbres and dynamic levels. As depicted in Figure 1,therearestopknobsavailableneartheperformerthat allow to select the pipes where the air will flow through. Each stop controls a specific rank of pipes, corresponding with the keys of the keyboard, that realizes a particular sound quality, ranging from flutes and strings to reeds and diapasons. By skillfully choosing and blending stops, the organist can imitate the sounds of an orchestra, produce unique tonal effects, and adapt the instrument’s sound to suit different musical styles and acoustical environments. 3https://www.organroxx.com Proc. of the 17th International Symposium on CMMR, London, UK, Nov. 3-7, 2025 299 Exploring the Classification of Pipe Organ Sounds Fig. 1. Console of an organ, with three manuals (pedal board for the feet not shown), and a range of stop knobs that are used to select the mixture of timbres. Registration is crucial for expressive playing, as it allows the organist to shape the music’s emotional and dynamic contours. Mastery of registration techniques is essential for interpreting organ repertoire authentically and creatively, making it a key aspect of organ performance and artistry. For the major part of historical organ traditions, the exact choice of registration is delegated to the performer. There are historical sources providing general directions, and there is a rich tradition of conventions, but for most individual compositions it is not prescribed exactly which stops to select. An exception to this is the era of the French Baroque, in which the title of the movement or composition often indicates the registration. For example, a piece or movement with the title ‘Grand Jeu’ always should be played with reeds and cornets, or a piece called ‘Tierce en taille’ always features a solo melody in the tenor with the Tierce stop, with a softer accompaniment played on a different keyboard and a flute sound for the bass. This correspondence between title and timbre for French Baroque organ music allows us to assemble a collection of ground-truth labels for registration recognition. 3Data To train our various models, we assemble different selections of tracks from the OrganRoxx collection, with different sets of labels. For organ type recognition, we use expert-provided labels of organ types for 139 organs that are present in the OrganRoxx collection. The labels include: ‘Renaissance’, ‘Baroque’, ‘Romantic’, and ‘Symphonic’. For each of these organs, Proc. of the 17th International Symposium on CMMR, London, UK, Nov. 3-7, 2025 300 V. During et al. we retrieve all tracks that are performed on that organ and include those in the respective classes. For the recognition of individual organs, we take all those organs that have more than 100 tracks available in the collection. This results in a set with 239 instruments and c. 56 thousand tracks. For the registration recognition task, we first compile a list of composers who wrote music for the organ during the French Baroque period.4For each of these composers, we retrieve all their compositions from the OrganRoxx collection and we select those with a title matching one of the following classes: Plein Jeu,Grand Jeu,Fond d’Orgue,Flûtes,Duo,Récit de Cornet,Récit de Nazard, Récit de Cromhorne,Tierce en Taille,Basse de Cromorne,Basse de Trompette, Voix Humaine,andGrand Choeur.5Next, we do a manual check to exclude the obvious mislabeled tracks. These include performances in which the organist did not obey the registration (e.g., by playing a ‘Basse de Trompette’ with the Cromorne, or a ‘Tierce en Taille’ without the Tierce), tracks that feature multiple registrations (e.g., a ‘Basse de Trompette’ that also features sections with a Cornet solo), or performances on organs that were not designed for the French tradition (e.g., a nineteenth century Dutch organ). The resulting set contains c. two thousand tracks. For the intra-track registration recognition, we use the same data set as for the classification of full tracks. Since we do not have a proper ground truth for this task, we select a few tracks and we examine the predictions of our classifier in detail. These include tracks that contain multiple movements (using different registrations), or that feature different registrations within one movement (e.g.,. a ‘dialogue’ between Trompette and Cornet). 4Methods 4.1 BYOL-A Our approach builds on the BYOL-A framework [7], a self-supervised method for learning audio representations without labeled data. Inspired by the BYOL architecture from computer vision, BYOL-A uses two networks, an online and atargetnetwork,thatprocessdifferentaugmentationsofthesameaudioinput. The online network learns to predict the target network’s output, which is updated via a moving average. Unlike contrastive methods, BYOL-A does not rely on negative samples. Audio is first converted to log-mel spectrograms and then passed through a convolutional encoder to produce 2048-dimensional embeddings. These embeddings are robust to augmentations like mixup and random resize crop. We use the AudioNTT2020 pretrained model as provided in BYOLA[7], which was trained on datasets containing a wide range of real-world audio 4https://en.wikipedia.org/wiki/French_organ_school 5Although Grand Choeur is not strictly a French Baroque registration, it complements the set of timbres. We didn’t include Trio because of the large variability of registrations within the collection. Proc. of the 17th International Symposium on CMMR, London, UK, Nov. 3-7, 2025 301 Exploring the Classification of Pipe Organ Sounds Table 1. Classification report on the test set for the organ type task, including precision, recall, F1-score, and support per class. Label Prec. Rec. F1 Sup. Label Prec. Rec. F1 Sup. Baroque 0.98 0.98 0.98 2550 Romantic 0.94 0.96 0.95 1053 Renaissance 0.94 0.91 0.93 110 Symphonic 0.91 0.64 0.75 112 recordings (such as AudioSet). This contributes significantly to the generalization ability of the learned embeddings. We use the pretrained model without fine-tuning and train classifiers on top of the extracted embeddings. 4.2 Data Processing and Embedding All audio files are resampled and normalized to conform to BYOL-A’s input specifications. For the slicing-based task, the audio is segmented using a 5-second fixed-length window with a 1-second hop size. Each resulting slice or full track is then converted into a 2048-dimensional BYOL-A embedding. 4.3 Model Training We train a multi-layer perceptron (MLP) classifier for each task, using a shared architecture adapted to the input dimensionality and number of output classes. The architecture consists of three fully connected hidden layers (488, 334, and 179 neurons, respectively), each followed by Batch Normalization, ReLU activation, and dropout (rate = 0.30) for regularization. The output layer maps the BYOL-A embeddings to class scores. Hyperparameters were optimized using the Optuna framework. We train the models for up to 25 epochs using the Adam optimizer with an initial learning rate of 0.000723, which is adjusted using a StepLR scheduler. Early stopping is applied to prevent overfitting. Performance is evaluated on both training and validation sets using accuracy, macro F1-score, and confusion matrices. 5Results 5.1 Organ Type Recognition Table 1 shows the performance of the classifier on identifying organ types. The overall accuracy is high, at 0.964, indicating that the model can distinguish the organ types reliably. Baroque and Romantic types are classified with particularly high precision and recall. Misclassifications mostly occur between Renaissance and other types, and between Symphonic and Romantic, which is understandable given overlapping musical characteristics in instrumentation and harmonic language. Proc. of the 17th International Symposium on CMMR, London, UK, Nov. 3-7, 2025 302 V. During et al. Table 2. Classification report for the top-10 organs by validation support, including precision, recall, and F1-score. Organ Prec. Rec. F1 Organ Prec. Rec. F1 Collegio N. Mazza, Padova 0.91 0.98 0.94 Groote Kerk, Maassluis 0.77 0.89 0.82 Laurenskerk, Alkmaar 0.78 0.85 0.81 Martinikerk, Groningen 0.80 0.79 0.79 Saint-Eustache, Paris 0.76 0.80 0.78 Saint-Ouen, Rouen 0.85 0.86 0.85 Sankt Jacobi, Hamburg 0.73 0.83 0.78 Saint-Sulpice, Paris 0.88 0.82 0.85 Saint-Martin, Dudelange 0.78 0.87 0.83 Sainte-Madeleine, Paris 0.82 0.84 0.83 5.2 Organ Identification For the organ identification task, Table 2 shows the precision, recall, and F1-score for the ten organs with the highest number of validation examples. The F1-scores for these classes range from 0.776 to 0.942. The highest score is observed for the organ at Collegio N. Mazza in Padova, with a precision of 0.910 and a recall of 0.976. The lowest F1-score among the top-10 is found for Sankt Jacobi, Hamburg, with a precision of 0.726 and a recall of 0.833. Over 20 independent training runs, the model achieved a mean final accuracy of 0.8103 with a standard deviation of 0.0024 for all 239 classes. 5.3 Full Track Registration Recognition Figure 2 shows a t-SNE embedding of the BYOL-A embedding vectors for the dataset for the registration recognition task. Note, that this is prior to our classifier. The various registration classes clearly form clusters in this space. Table 3 shows the performance of the MLP on classifying the BYOL-A embedding vectors. This appears very successful. Half of the classes even reaches perfect results. Inspection of the confusion between the classes shows that the misclassifications are understandable. Notably, there is some confusion between Grand Choeur and Grand Jeu, which indeed share timbre similarities. Table 3. Classification report on the test set for the the Registration Recognition task with precision, recall, F1-score, and support for each class. Class Prec. Rec. F1 Sup. Class Prec. Rec. F1 Sup. BasseCromorne 0.94 1.00 0.97 15 GrandJeu 0.96 0.96 0.96 70 BasseTrompette 0.96 0.93 0.95 29 PleinJeu 1.00 1.00 1.00 93 CromorneEnTaille 1.00 1.00 1.00 25 RecitCornet 1.00 1.00 1.00 19 Duo 0.95 1.00 0.98 21 RecitCromorne 1.00 1.00 1.00 10 Flutes 1.00 0.89 0.94 19 RecitNazard 0.85 1.00 0.92 11 FonddOrgue 1.00 1.00 1.00 7 TierceEnTaille 1.00 1.00 1.00 43 GrandChoeur 0.88 0.83 0.86 18 VoixHumaine 1.00 1.00 1.00 26 Proc. of the 17th International Symposium on CMMR, London, UK, Nov. 3-7, 2025 303 Exploring the Classification of Pipe Organ Sounds Fig. 2. t-SNE embedding of the BYOL-A embeddings. Each point represents a CDtrack. Colors indicate the class. 5.4 Intra-Track Registration Recognition Since a full evaluation of the task of Intra-Track Registration Recognition is not possible because of the lack of proper ground-truth. We inspected a number of hand-picked tracks by comparing the predicted labels for the time slices with the audible contents of the audio track. We now briefly report the results for one track that seems representative of the kinds of confusion we observed. This track contains six separate movements:61. Grand Jeu, 2. foundational stops, but not with the typical slow-moving texture of a ‘Fond d’Orgue’, 3. Voix Humaine for both hands, 4. Récit de Cornet, 5. Flûtes, and 6. Grand Jeu. The predictions of our model show interesting confusion. Slightly less than half of the slices for the first movement are correctly classified as Grand Jeu, the other half as the closely related class Grand Choeur. The second movement is largely classified as Duo, while Fond d’Orgue would have been better according to the timbre. Apparently, our classifier has learned the structure of the music (having two voices) rather than the timbre for the Duo class, as this movement has two voices indeed, but lacks the characteristic Cornet sound of the Duo. Surprisingly, a few slices are classified as Basse de Cromorne, which is a very distant class. There is nothing perceivable in the audio that would justify this mistake. The third movement shows substantial confusion between Basse de Cromorne and Grand Jeu. This is not entirely surprising since the Voix Humaine and the Cromorne have a somewhat similar timbre. Only one of the slices has been correctly classified as Voix Humaine, which is remarkable since the movement as 6ArecordingofNoel en Grand Choeur by J.J. Beauvarlet-Charpentier. Proc. of the 17th International Symposium on CMMR, London, UK, Nov. 3-7, 2025 304 V. During et al. a whole has been classified correctly. We speculate that this is caused by the classifier recognizing the typical structure of a Voix Humaine movement, rather than the distinctive timbre of the Voix Humaine pipes. The fourth movement shows substantial confusion between Duo and Récit de Cornet. This movement mainly has two voices indeed, and features the Cornet sound which is typical for the Duo, however it lacks the Tierce in the bass. The fifth movement is correctly predicted as Flûtes for all slices. The final movement is largely classified correctly, with minor confusion with Grand Choeur. From this example, and the others we examined, it becomes clear that the intra-track classification is challenging and shows a lot of confusion, while the classification of the full track is mostly correct. 6ConcludingRemarksandFutureWork In this short study, we showed that the Bootstrap Your Own Latent for Audio (BYOL-A) framework provides a useful foundation for various tasks concerning classification of pipe organ sounds. For the organ type recognition task we obtained a very good performance. The recognition of individual organs shows very impressive performance as well, especially considering that it is a 239-class problem. The registration recognition task almost reaches perfect performance for full tracks. The intra-track registration classification task we only evaluated by observing a few examples. It became evident that the error rate for classifying audio slices is much higher than for classifying full tracks. In certain instances, classes that are mostly correctly predicted for full tracks, are not correctly predicted for time slices. This could be caused by the diversity of timbres that is present in the full tracks that were used in the training set, or by the model learning structural or textural features instead of timbre. There are many directions to explore in future work. We plan an effort to annotate intra-track registrations. That will provide a better ground-truth for the registration classification task. We also will explore the performance of our methods on organ music from other traditions than the French Baroque, and we continue to work on the organ type classification task with the aim to improve accuracy and eventually also to include more fine-grained classes. On https://github.com/pvankranenburg/cmmr2025,weprovidecodeandfurther details. Acknowledgments. We are grateful to OrganRoxx for making their audio collection available for this research project. Disclosure of Interests. The authors have no competing interests to declare that are relevant to the content of this article. Proc. of the 17th International Symposium on CMMR, London, UK, Nov. 3-7, 2025 305 Exploring the Classification of Pipe Organ Sounds References 1. Ciaramella, A., Lauro, E., Martino, S., Falanga, M., Tagliaferri, R.: Modeling and generating organ pipes self-sustained tones by using ica. Journal of Signal and Information Processing 2(3), 141–151 (2011). https://doi.org/10.4236/jsip.2011. 23018 2. Comerford, P.: Simulating an organ with additive synthesis. Computer Music Journal 17(2), 55–65 (1993), http://www.jstor.org/stable/3680869 3. Dumenil, P.: Predictive Analytics for Pipe Organ Registrations. Master’s thesis, Sorbonne Université, Paris (2024) 4. Fischer, J.L.: Shock wave characteristics in the initial transient of an organ pipe. In: Bader, R. (ed.) Computational Phonogram Archiving, pp. 269– 304. Springer International Publishing, Cham (2019). https://doi.org/10.1007/ 978-3-030-02695-0_13 5. Hamadicharef, B.: Artificial Intelligence-Based Approach to Modelling of Pipe Organs. Thesis, University of Plymouth (2005), https://pearl.plymouth.ac.uk/ secam-theses/120 6. Miklós, A., Angster, J., Pitsch, S., Rossing, T.D.: Interaction of reed and resonator by sound generation in a reed organ pipe. The Journal of the Acoustical Society of America 119(5), 3121–3129 (2006) 7. Niizumi, D., Takeuchi, D., Ohishi, Y., Harada, N., Kashino, K.: Byol for audio: Self-supervised learning for general-purpose audio representation (2021), https: //arxiv.org/abs/2103.06695 8. Sawicki, J., Abel, M., Schöll, E.: Synchronization of organ pipes. The European Physical Journal B 91(2), 24 (feb 2018). https://doi.org/10.1140/epjb/ e2017-80485-8 9. Schedl, M., Gómez, E., Urbano, J.: Music information retrieval: Recent developments and applications. Foundations and Trends in Information Retrieval 8(2-3), 127–261 (2014). https://doi.org/10.1561/1500000042 Proc. of the 17th International Symposium on CMMR, London, UK, Nov. 3-7, 2025 306