A Chant Scholar and an AI Agent: Building Plainchant OMR from Legacy Ground Truth
Robert Klugseder, Vienna
TL;DR
This is an experience report. A chant scholar without training in machine learning worked with an AI coding agent (Claude in Claude Code) to build optical music recognition (OMR) for medieval plainchant and to check training data that student assistants had transcribed by hand in Transkribus. The scholar decided all questions of chant, notation and acceptance. The agent proposed and implemented models, tools and measurements. The recognition model uses two ideas from speech and handwriting recognition, a CTC head and an attention decoder trained and scored together, and a language model of chant melodies from the Cantus Database. A third idea comes from chant scholarship: comparison with other manuscripts’ versions of the same chants. On a fixed test set of 3,084 lines, the symbol error rate fell from 7.09 % to 5.71 %. To check the legacy data, every page was read by a model that had not been trained on it, and only disagreements went to the scholar; from the eighth source on, a multimodal language model judged each disagreement first. With the checked data from 27 sources, 27 manuscripts were transcribed in full (99,404 staff systems, most of them unchecked machine readings). The scholar’s random samples of unseen pages found 0.3–2.3 % errors in notes and clefs. The report describes how the work was divided, which working habits helped, and where they failed.
1 Introduction
I am a musicologist who works on medieval plainchant. I am not a computer scientist. In October 2026, working with an AI coding agent, I built a system that reads the notation of chant manuscripts and a procedure that checks older, hand-made training data. This report describes how that worked: what I contributed, what the agent contributed, and which working habits other researchers in the humanities, and perhaps machine-learning researchers who work with domain experts, can reuse.
Two questions drove the work. The first is one a chant scholar asks naturally. Most of an antiphoner or gradual belongs to a common repertoire. Before I open a new manuscript, I already know most of what it contains, and the Cantus Database [2] holds the texts of tens of thousands of chants, many with their melodies in the Volpiano encoding [3]. Can this knowledge help a machine read the notation, as it helps me? The second question is practical. Older ground truth for OMR, that is, hand-made transcriptions for training and testing a model, exists in Transkribus [4]. Student assistants created it: 47 books with notation on 1,256 pages. It contains typing errors, changes of convention within a book and some systematic misreadings. It has to be checked before training, and checking every line by hand is too expensive.
Commentators on AI-assisted programming have observed that writing code has become cheap, while human attention has not [5, 6, 7]. My experience confirms both halves. The code, the training runs and the measurements in this project cost little and came fast; a scholar with a basic technical background can now direct demanding machine-learning work. What stayed expensive was my attention as the expert. Without it, in setting the conventions, calibrating the judge and checking samples of the results, none of the output would have been usable.
This report continues an earlier one [1], which turned Volpiano transcriptions from the Cantus Database into about 30,000 lines of training data. Here I describe the next step: a reading model, a review cycle for the legacy ground truth, and the transcription of 27 manuscripts.
I have to be clear about one limitation from the start. The agent designed and implemented the models, chose the training settings and ran all measurements. I can judge whether a transcription matches the image, and I decided every question of notation. I cannot verify the inner workings of the networks. I therefore describe the technical side only at the level at which I can take responsibility for it, and I say where a number rests on the agent’s measurement alone.
2 Sources and starting point
Chant survives in many notational families. The manuscripts in this project are written in square notation, in Gothic notation of the German type (“Hufnagel”), in Bohemian Gothic notation and in Klosterneuburg notation. Some later books mix square notation with mensural notation for hymns. The older data from the first report also contain French, Messine and Cistercian notation. The manuscripts span five centuries; the printed books extend this to seven.
The model does not output pitches. It outputs the position of each note on the staff (on a line or in a space, counted from the bottom) and the clef as written. Pitches follow from the clef. This is how the legacy ground truth was written, and it keeps the reading of the image apart from its musical interpretation. Division marks and section bars are encoded; liquescent note forms and custodes (guide notes at the end of a line) are not.
Two bodies of training data were available (Table 1). The first comes from the earlier report: 32,746 line images from ten sources (eleven exports), aligned with their Cantus Database transcriptions. Its fixed test set of 3,084 lines never changed during the project. The second is the legacy Transkribus ground truth. Three student assistants, here called annotators A, B and C, wrote it with a keyboard, token by token. Writing staff positions by hand in this way is tedious, and errors are to be expected. Each annotator also developed conventions of their own, for example in which order the two notes of a pes are written, or whether a flat sign stands where it is in the image or before the note it affects.[¹] This is no reflection on the annotators’ care; such errors are normal in a tedious manual task.
Material Lines or systems Notation Volpiano-based ground truth from the first report, extended (ten sources) 32,746 lines various, incl. French, Messine, Cistercian — of which fixed test set 3,084 lines Legacy Transkribus ground truth checked here (27 sources) 7,400 training lines 14 square (two with mensural hymns), 12 German Gothic, 1 Bohemian — of which twelve books of the Franciscan library in Graz square Manuscripts transcribed in full (27) 99,404 systems square and Gothic Printed chant books (12, kept separate) 29,670 systems square
3 Division of labour
What I did. I chose the sources and the order in which they were processed. I decided all questions of notation and transcription. Some examples: a flat sign applies to the next note on the B position, even when the scribe wrote it earlier; a sharp sign on the B position is read as a natural; a key signature at the start of a line is written directly after the clef; an extended reciting bar counts as one note; mensural hymns are excluded from training and only flagged in the transcription. I recognised the systematic errors in the legacy data that the measurements pointed to, for instance that one annotator wrote a special Bohemian climacus form with two instead of three notes throughout a whole antiphoner. I had the idea of letting an agent pre-check the disputed lines. I decided every remaining disputed case on review pages, corrected random samples of the machine transcriptions, and approved the training of each general model.
What the agent did. The agent wrote the project code: the conversion of the Transkribus exports, the models, training and decoding, the measurements and the review pages. It built on the training tools of the browser-based OMR application for late-medieval plainchant that Jorge Calvo-Zaragoza and his team at the Pattern Recognition and Artificial Intelligence Group (PRAIG) of the University of Alicante developed together with me [8], and on the model from the first report. It ran the experiments, kept a project diary and drafted the reports in German; I revised them. It also proposed most of the technical steps, for example having each page of the legacy data read by a model trained without it, and it built and tested the judge.
Working rules. A handful of rules, most of them introduced after something went wrong, made this division workable:
- A diary with sources. Every number goes into a diary together with the file it comes from. Agent sessions end, and long sessions are summarised automatically; the diary is the project’s memory.
- Review pages instead of tables. Each system appears with its image, the competing readings set in the Volpiano font, an editable correction field and buttons. My decisions are stored with the page, and the agent reads them back. I could review remotely.
- Small, editable samples. Samples have 20–40 systems and an editable correction for each one. An early sample that only asked “right or wrong” showed no patterns.
- Expert questions go to the expert. The agent marks questions of chant or notation and leaves them open instead of answering them itself.
- Training only after review, and only with explicit approval for the general models.
- Independent checks of the agent’s work. Before this report was finished, five separate agent instances checked the German draft against the project files. They found 22 factual errors. One came from the agent’s own diary: it had called the last model the best one, although two earlier models scored slightly better on the fixed test set.
4 The reading model in brief
The agent built the model from parts that are well established in speech and handwriting recognition. I describe them only as far as I can explain what they do; the architecture, the settings and the decoding weights were chosen by the agent following published work.
Two heads, read together. A convolutional and recurrent encoder (CRNN) turns the line image into a sequence of features, as in earlier work on music recognition by Calvo-Zaragoza and colleagues [9, 10]. Two heads read this sequence. A CTC head (connectionist temporal classification, a standard way of reading a sequence without knowing where each symbol starts) assigns a symbol, or a blank, to each step along the encoded line. Through the recurrent encoder it sees the whole line, but it does not condition each symbol on the symbols it has already written. An attention decoder writes the symbol sequence one token at a time and can use the tokens it has already written. Both heads are trained together, and when reading, the decoder’s candidates are scored by both. Joint CTC/attention training was introduced in speech recognition [15], and joint decoding combines the two scores during the search [16]. Wick, Zöllner and Grüning applied this combination to the recognition of text lines [17]. The AMNLT study of chant recognition compares CTC-based systems with autoregressive attention-based models [11], and Kurek and Hajič use an attention-based model [14]; neither reports joint CTC/attention decoding. On the fixed test set, the symbol error rate (the share of symbols that are wrong, missing or added) was 9.56 % for the model of the first report and 7.09 % after retraining it with CTC alone. Joint training and scoring brought it to 5.71 %.
A language model of chant melodies. From the Cantus Database the agent extracted about 193,000 melody lines (184,000 for training) and converted them into the same staff-position tokens. A small transformer learned to predict the next note. Chants on the folios used for validation and testing were excluded. The melody model is used in two ways. The decoder starts its training from what the melody model has learned. And when reading, the melody model’s score for each candidate is added to the scores of the two heads, a method known as shallow fusion [18]. Earlier work by the Würzburg group used fully convolutional models for square notation [12]. Hartelt and Puppe later reported a preliminary attempt to correct pitch-position errors with an n-gram model of note pitches; it introduced about as many errors as it corrected [13, p. 15]. Our model differs in size, in training data, and in that it chooses between candidates instead of correcting the result.
Versions of the same chant. For each line, the folio inventory in the Cantus Database names the chants on the page. Their versions in other manuscripts are collected, and each candidate reading gets a bonus for its similarity to them. Other sources are often notated at a different pitch level. A candidate with a different clef from the one the model read therefore gets this bonus only if it moves the line into the transposition on which most versions agree. Using the versions directly as candidates did not work: they bring their own variants with them, and on an unseen manuscript the best borrowed version was worse than the image reading in 680 lines and better in only 285. Other versions help to decide between readings, not to copy them.
Where this helps. On the familiar sources of the fixed test set, the repertoire knowledge changed almost nothing. It helps on a manuscript the system has never seen. For this test the agent trained models without the twelfth-century antiphoner A-KN 1012 from Klosterneuburg; the melody model also left out this manuscript’s chants. Other manuscripts’ versions of the same chants remained available, which is intended: the test measures a new witness of a known repertoire, not unknown melodies. Only 2.0 % of the chants of A-KN 1012 have no version with melody elsewhere in the database. Joint scoring lowered the symbol error rate from 15.69 % (CTC alone) to 13.03 %. With a wider search and the repertoire knowledge, it fell from 13.08 % (the same search without repertoire knowledge) to 12.63 %. The pitch error rate, that is, the errors in the melody after the clefs are applied, fell from 18.47 % to 15.52 %. Most of this gain comes from corrected clefs. The repertoire helps above all to decide at which pitch level a line lies.
5 Checking the legacy ground truth
Figure 1 shows the review cycle. The agent first harmonised the annotators’ conventions according to rules I set. It then split the ground-truth pages of a source into five parts, trained a model on four parts and let it read the fifth, repeating this for each part (five-fold cross-validation). So every page was read by a model that had not been trained on it. Only lines where the reading and the ground truth disagreed in notes or clefs went on to the next step.
Figure 1: The review cycle. Thick boxes are the expert’s work; the dashed box is the multimodal judge; all other steps were built and run by the agent.
Who was right? For the first seven sources I decided every disagreement myself by looking at the line image. The pattern was clear. Where pitches differed, the model was usually right; the ground truth contained typing errors, many of them on neighbouring keys of a German keyboard. Where the model had read fewer notes than the ground truth, the ground truth was usually right. The model tends to drop the second of two equal notes, and it often overlooks flat signs. Where clefs differed, the ground truth was usually right. In A-Wn 15484, for example, F clefs near the gutter are partly hidden and look like C clefs. The cycle also found three systematic errors. On the first 66 pages of one gradual, annotator B wrote the two notes of a pes from top to bottom, so that each pes looked like a clivis (397 swapped neighbouring notes); the same annotator did this in three of the Graz books. In one Bohemian antiphoner, a special climacus form was written with two notes instead of three throughout. And typing errors on neighbouring keys occurred in all sources of one annotator. Similar label errors are known from other benchmarks [21], and using a model’s disagreement to find them has been described for handwritten text [22].
Repairing a convention from the image. A rule cannot repair the pes order, because a real clivis looks the same in the transcription. Instead, the model scored each falling pair of notes in a neume group against the image, once as written and once swapped, and kept the better version. The pitches stayed those of the ground truth. On pages without this error, the procedure changed only one line in 176. I checked a random sample of twelve corrected systems and found no error.
A judge before the expert. Deciding disagreements was the most time-consuming step for me. I therefore suggested letting an agent check them first, and the agent tested whether a multimodal language model could do this, an idea related to “LLM as a judge” in text evaluation [23]. A small exploratory study found that the multimodal models tested in 2024 struggled to transcribe selected excerpts of modern scores [24]. More recent work shows that supplying an OMR transcription alongside a score image can improve performance on music-understanding tasks [25]. These results motivated giving the judge candidate transcriptions as context. They do not show how reliably a model can decide between competing readings of medieval notation; that I tested separately against my own decisions. Our judge never transcribes freely. It sees the line image, the ground truth, the model’s reading and the positions where they differ. It chooses one of them or writes its own version, with a confidence level and a short reason.
How the judge reads. The judge is the same kind of model as the agent, started as a separate sub-agent for each batch of 10–15 systems; ten of them run in parallel. It was not trained on chant. What it knows about neumes comes from the general training of the language model. For the task it receives an instruction of about 600 words, much as one would brief a new student assistant. The agent wrote the instruction; the rules specific to this project in it, such as the treatment of flat signs, come from me. It explains the staff-position tokens, that a pes is read from bottom to top and a clivis from top to bottom, that a flat sign is written directly before the note it affects, and that the custos is not transcribed. The judge is instructed to read only its task file and the line images, and my decisions are not part of its task. For each system it has to decide every point of disagreement from the image and state what it sees there.
How the image is shown mattered as much as the instruction. When the judge received the line in one or two large pieces, which the image tool reduced to preview size, it agreed with me in 20 of 30 systems. When the line was cut into overlapping pieces of at most 600 pixels and enlarged twice, it agreed in 25 of 30; among systems where the readings differed in pitch, it was right in 10 of 12 instead of 7. Two of its reasons, translated from German (both verdicts match my decisions):
“In the second piece there are two neumes of the same shape, like an m. The heads of both lie on the top line, the last leg ends on line 3. The first is therefore l4 l4 l3, not l3 l3 l2.”
“Behind the flat sign stands a three-note scandicus, rising diagonally: a lozenge on line 2, a lozenge in the space above it, a head on line 3. The lozenge on line 2 is real; the flat sign does not replace it.”
Some things remain hard even when enlarged: dense runs of lozenges, positions at ledger lines and the feet of a pes close to a stem. When it is unsure, the judge tends to confirm the ground truth. A check costs several thousand tokens per system; our counts on printed books range from about 5,000 to 9,000.
In a blind test on 129 disputed systems of one gradual, which I had decided before, the first judge (Claude Fable 5.1) agreed with me in 97 cases (75 %), and in 43 of 47 cases (91 %) where it reported high confidence. I then examined its 32 disagreements. After 20 of them I stopped, because in all 20 my version was right. Two misreadings of the notation explained most of them. Where a flat sign stood before a pes, the judge took the first note of the pes for the sign and deleted it. And it looked for the middle note of a scandicus directly above the first note instead of to the upper right. These two rules were added to the judge’s instruction. With them, and with the less expensive model Claude Opus 5.5, the judge decided 44 of 52 selected cases correctly (85 %): the 32 earlier disagreements plus 20 cases it had judged correctly before. Because the model and the instructions changed at the same time, I cannot say which of the two helped. In later use across seven sources, high-confidence verdicts agreed with mine in 54 of 63 cases (86 %).
I also read the judge’s reasons. They describe the neume forms correctly and use real knowledge of chant notation, which I had not expected. They helped me even where the verdict was wrong, because they named the neume that had been misread.
Get Robert Klugseder’s stories in your inbox
Join Medium for free to get updates from this writer.
Accepting a known risk. I then adopted a simple rule. If the judge agreed with either the ground truth or the reading, the system counted as checked and no longer appeared on my review page. In the first three sources this removed 70 % of the systems (234 of 333). This rule saves work; it does not validate anything. The agent measured its cost on 80 systems I had already decided before the rule hid them: in 15 of them (19 %) my decision differed from the agreed version, almost always by one note. I accept this residual risk, because otherwise the review would not have been possible at this scale. For the twelve books of the Franciscan library in Graz I then saw only 3 to 47 systems per book.
6 Scaling up
Forgetting and replay. When the agent fine-tuned the reading model on the checked Transkribus lines alone, the model forgot the older sources: its error on the fixed test set rose from 5.71 % to 10.51 %. Machine-learning research calls this catastrophic forgetting [19]. We noticed it only when the agent measured these models on the fixed test set afterwards. Mixing the old ground truth back into training, a method known as replay [20], removed the problem at once (5.66 %). Giving the new lines three times the weight kept the error on the new sources’ test pages as low as that of the model trained on them alone, while the fixed test set stayed at 5.64 %. Since then, every general model has also been measured on the unchanged fixed test set.
One model for all sources. The final general model was trained on the old ground truth and the checked lines of all 27 Transkribus sources. Its error rate on the fixed test set is 5.68 %, slightly higher than that of the first two replay models. On the test pages of the thirteen most recently checked sources (256 lines), it makes 2.55 % errors in notes and clefs (95 % bootstrap interval 1.99–3.18 %). In an earlier measurement on nearly the same lines, a general model that had not seen these sources made 9.26 %.
Transcribing the manuscripts. The second-to-last general model, which differs little from the final one, read all unread systems of 25 manuscripts: 82,331 systems in about one hour on a single GPU. Together with the ground truth and the checked systems, and with a sister volume that had no ground truth of its own, there are now complete transcriptions of 27 manuscripts with 99,404 staff systems, including a gradual with five-line staves that was added after the first version of this report. More than nine in ten of these systems are machine readings that no one has checked. To measure their quality, I corrected random samples of 30 systems from unseen pages in five books. The error rates in notes and clefs were 0.31, 0.83, 1.13, 1.25 and 2.30 %. Samples of this size show the order of magnitude for these five books; they are not an accuracy figure for all 99,404 systems. The most frequent errors were overlooked flat signs, confusion of C and F clefs, and missing key signatures. In the book with the highest error rate, judges based on Claude Opus 5.5 then checked the less confident half of all systems for clefs and key signatures only. They proposed 440 corrections, which the agent applied without my review, and left 30 cases for me to decide.
Printed books. Twelve printed chant books from the same Graz library (29,670 systems) form a separate project with separate models, because their notation and image sizes differ. Legacy ground truth exists for only five of them, and I decided not to proofread them. Here the judge took the expert’s place: it examined the least confident tenth of the systems in each book, in some books more, and its confident verdicts became training lines for the next model. In one two-colour print, the judges’ corrections revealed that the black notes were printed up to one staff position (from a line to the adjacent space) below the red staff lines; a script now measures and corrects this shift. I have no measured error rate for the printed books as a whole.
7 Patterns, and what I cannot check
These are the patterns I would reuse in other fields where an expert has to check large amounts of machine output.
Let the model read blind, and show the expert only disagreements. Cross-validated reading turns a pile of legacy data into a short list of questions. Taken together, the answers also reveal more: the pattern of who was right exposed systematic errors that no one had noticed.
Put a judge before the expert, but measure it against the expert first. A blind test on cases the expert has already decided shows what the judge can do. The judge improves when the expert analyses its errors. Only then can the expert decide how much risk to accept for which data, instead of trusting a confidence score.
Use the knowledge of the domain where it is strongest. For chant, that is the repertoire. It helped most where the image was hardest to interpret, at the level of clefs and transposition, and on an unseen manuscript. Whether such background knowledge helps has to be measured on material the model has not seen.
Keep a test set that never changes. The fixed test set exposed forgetting that the test pages of the new sources did not show.
Make the expert’s decisions cheap and durable. Review pages with images, editable fields and stored decisions; small samples; a diary that records every number with its source.
Check the agent’s reports independently. Agents write fluent, confident text. Independent agent instances, given only the draft and the files, found errors that the agent had missed.
Spend the expert’s attention where it counts. In this project, code and training were cheap; the expert’s time was the scarce resource. Each pattern above serves to put that time into calibration and quality control rather than into routine checking.
There are clear limits. I cannot verify the architecture or the training of the models; I can only judge their output on images. Test pages per source are small, between 9 and 57 lines, so differences of a few notes between two models are within noise. The transcriptions are mostly unchecked machine readings; the samples measure their quality but do not correct them. The test on an unseen manuscript measures a known repertoire, not unknown melodies. One manuscript failed at first: a gradual with five-line staves was read with 45.5 % errors in note positions. The cause was geometry, not notation. The crop assumed four staff lines instead of five, and one clef symbol was missing from the model’s vocabulary. After both were corrected, the model read the manuscript’s test lines with 2.39 % errors, and all 5,923 systems were transcribed. The judges first read many F clefs as C clefs, until I pointed out the shape that tells them apart: the C is open on the right, the F is closed by its loop. After that they were right in 18 of 19 cases. The judge’s agreement rule leaves a residual risk of about one system in five among those it removed from review. And because no code, data or models are released (Section 8), others cannot reproduce the numbers.
8 Disclosure and availability
I worked with an AI coding agent: Claude Opus 5.5 and Claude Fable 5.1, both in Claude Code. The agent wrote the project code and ran all training on a single workstation GPU. It built the review pages, kept the project diary and wrote the German draft of this report, which I revised. The English text was drafted by the agent from the revised German draft and edited by me. I selected the sources, made every decision on chant and notation, decided the disputed cases, and am responsible for the content of this report.
I publish this report as a description of a working method that others can adopt. I do not publish the code, the models or the checked ground truth, because I would only release what I fully understand. The figures in this report come from the project diary and the measurement logs, which record the file each number comes from.
Acknowledgements
I thank Jorge Calvo-Zaragoza and his team at PRAIG, University of Alicante, for our joint work on OMR for chant [8], on which this project builds. I thank the three student assistants whose transcriptions made this work possible.
References
[1] R. Klugseder: From Volpiano to Pixels: 30,000 Lines of OMR Ground Truth for Medieval Chant. arXiv:2610.16862v1 [cs.CV], 2026.
[2] Cantus Database. Directed by D. Lacoste (2011–present), T. Bailey (1997–2010), R. Steiner (1987–1996); web development J. Koláček (2011–2023), DDMAL (2023–present). https://cantusdatabase.org (accessed 22 October 2026)
[3] K. Helsen, D. Lacoste: A report on the encoding of melodic incipits in the CANTUS database with the music font “Volpiano”. Plainsong and Medieval Music 20(1), 2011, 51–65. https://doi.org/10.1017/S0961137110000197
[4] READ-COOP SCE: Transkribus. https://www.transkribus.org (accessed 22 October 2026)
[5] S. Willison: Writing code is cheap now. In: Agentic Engineering Patterns, 2026. https://simonwillison.net/guides/agentic-engineering-patterns/code-is-cheap/ (accessed 22 October 2026)
[6] M. Quimby: Code got cheap. Your attention didn’t. DEV Community, 21 October 2026. https://dev.to/max_quimby/code-got-cheap-your-attention-didnt-4fm3
[7] A. Salkever: Code is cheap. Proof is the bottleneck. Agentic AI Foundation blog, 8 September 2026. https://aaif.io/blog/code-is-cheap-proof-is-the-bottleneck
[8] J. Calvo-Zaragoza, J. Cerveto, R. Klugseder, A. Ríos-Vila, D. Rizo, A. Roselló: A lightweight browser-based optical music recognition tool for late-medieval plainchant sources. Unpublished manuscript, submitted to the Late-Breaking/Demo session of ISMIR 2026.
[9] J. Calvo-Zaragoza, D. Rizo: End-to-end neural optical music recognition of monophonic scores. Applied Sciences 8(4), 2018, art. 606. https://doi.org/10.3390/app8040606
[10] J. Calvo-Zaragoza, A. H. Toselli, E. Vidal: Handwritten music recognition for mensural notation with convolutional recurrent neural networks. Pattern Recognition Letters 128, 2019, 115–121. https://doi.org/10.1016/j.patrec.2019.08.021
[11] E. Fuentes-Martínez, A. Ríos-Vila, J. C. Martinez-Sevilla, D. Rizo, J. Calvo-Zaragoza: Aligned music notation and lyrics transcription. Pattern Recognition 170, 2026, art. 112094. https://doi.org/10.1016/j.patcog.2025.112094
[12] C. Wick, A. Hartelt, F. Puppe: Staff, symbol and melody detection of medieval manuscripts written in square notation using deep fully convolutional networks. Applied Sciences 9(13), 2019, art. 2646. https://doi.org/10.3390/app9132646
[13] A. Hartelt, F. Puppe: Optical medieval music recognition using background knowledge. Algorithms 15(7), 2022, art. 221. https://doi.org/10.3390/a15070221
[14] D. Kurek, J. Hajič jr.: Towards a foundational model for recognising diastematic Gregorian notation. arXiv:2606.31454v1, 2026.
[15] S. Kim, T. Hori, S. Watanabe: Joint CTC-attention based end-to-end speech recognition using multi-task learning. In: 2017 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2017, 4835–4839. https://doi.org/10.1109/ICASSP.2017.7953075
[16] T. Hori, S. Watanabe, J. R. Hershey: Joint CTC/attention decoding for end-to-end speech recognition. In: Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 2017, 518–529. https://doi.org/10.18653/v1/P17-1048; https://aclanthology.org/P17-1048/
[17] C. Wick, J. Zöllner, T. Grüning: Rescoring sequence-to-sequence models for text line recognition with CTC-prefixes. In: Document Analysis Systems (DAS 2022), LNCS 13237, Springer, 2022, 260–274. https://doi.org/10.1007/978-3-031-06555-2_18
[18] C. Gulcehre et al.: On using monolingual corpora in neural machine translation. arXiv:1503.03535v2, 2015.
[19] M. McCloskey, N. J. Cohen: Catastrophic interference in connectionist networks: the sequential learning problem. In: Psychology of Learning and Motivation, vol. 24, Academic Press, 1989, 109–165. https://doi.org/10.1016/S0079-7421(08)60536-8
[20] D. Rolnick, A. Ahuja, J. Schwarz, T. P. Lillicrap, G. Wayne: Experience replay for continual learning. In: Advances in Neural Information Processing Systems 32, 2019. arXiv:1811.11682v2
[21] C. G. Northcutt, A. Athalye, J. Mueller: Pervasive label errors in test sets destabilize machine learning benchmarks. In: Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks 1, 2021. arXiv:2103.14749v4
[22] S. Al-azzawi, E. Barney, G. Kovács, M. Liwicki: A human-in-the-loop label error detection framework applied to Arabic-script HTR datasets. arXiv:2601.16713v4, 2026.
[23] L. Zheng et al.: Judging LLM-as-a-judge with MT-Bench and Chatbot Arena. In: Advances in Neural Information Processing Systems 36, Datasets and Benchmarks Track, 2023, 46595–46623. arXiv:2306.05685v4
[24] J. Calvo-Zaragoza, E. Fuentes-Martínez, N. Luna-Barahona, A. Ríos-Vila: Can multimodal large language models read music score images? In: J. Calvo-Zaragoza, A. Pacha, E. Shatri (eds.), Proceedings of the 6th International Workshop on Reading Music Systems, 2024, 4–6. Proceedings volume: arXiv:2411.15741v1
[25] G. Yang, B. S. Zheng, V. Ebert, N. A. Smith: LEGATO 2: Toward multimodal sheet music recognition and understanding. arXiv:2607.05769v1, 2026.
