QuartSet: A String Quartet Dataset for Transcription and Source Separation of Real Instrument Recordings

QuartSet: A String Quartet Dataset for Transcription and Source Separation of Real Instrument Recordings Erika Rumbold[0009→0002→6637→5182] and George Tzanetakis[0000→0002→6844→7912] University of Victoria, Victoria BC, Canada {erumbold,gtzan}@uvic.ca Abstract. With the state-of-the-art for many MIR tasks being based on deep learning, it is crucial that there is a large amount of data to train these models. Many approaches opt for audio that has been synthesized from MIDI data, but these often lack the musicality and nuances of recorded musicians. We present QuartSet, a dataset of real instrument recordings and their accompanying scores for the tasks of audio-to-score transcription and score-informed source separation. The scores of QuartSet are provided in Kern notation, which is a simple format that is easily to manipulate. Instead of being simplified, as is common practice in automatic transcription research, the Kern scores contain every dynamic, articulation, and ornament marking as written by the original composers. This is so that the scores fully reflect what is present in the audio recordings. We show that QuartSet can be easily integrated with MIDI-synthesized data to create a larger, more diverse dataset that can be used to train a transcription model. When trained on both synthesized and real audio data, the model was able to produce better transcriptions of other real audio than the model that was trained only on synthesized data. Finally, we suggest potential methods for creating more audio and score data without synthesis. Keywords: audio-to-score transcription ·music source separation ·music information retrieval ·datasets ·chamber music 1Introduction Music Information Retrieval (MIR) is a field of research encompassing computational approaches to extract information from music. Within the scope of MIR are a variety of topics, including genre classification, music generation, source separation, and audio-to-score (A2S) transcription. For many MIR topics, the state-of-the-art algorithms utilize deep learning approaches, such as convolutional neural networks or transformer models to process musical recordings. Deep learning, by nature, requires a large amount of training data. However, this is a significant challenge for music research due to the lack of availability of data for training. In general, musical recordings can be di!cult to acquire due to copyright restrictions. Proc. of the 17th International Symposium on CMMR, London, UK, Nov. 3-7, 2025 582 E. Rumbold and G. Tzanetakis Many MIR tasks also require access to the isolated stems of the instruments or voices performed and recorded individually, which are not widely available. It is easier to obtain isolated stems for popular music because musicians are typically recorded separately with a metronome or backing track to synchronize the individual parts. Classical ensembles, however, are usually recorded together in order for the musicians to listen for and adjust to any fluctuations in tempo or intonation, or stylistic choices made by the rest of the group. This results in audio from other instruments that may “bleed” onto another instrument’s track, making the track unusable for deep learning-based MIR tasks. This lack of data is even more challenging for a task like audio to score (A2S) transcription, which has the additional requirement of musical scores, which are visual representations of what musical notes and events occur in the corresponding recordings. In this work, we present QuartSet, a collection of real string quartet recordings and their accompanying scores. These scores feature every dynamic, articulation, and ornament marking as written by the composer. The scores are provided in Kern notation, a simple format that is easy to manipulate. More details about the Kern score format are given in Section 3.2. In Section 2, we discuss current methods of dataset collection and their drawbacks, and give an overview of existing datasets for source separation and transcription of classical ensembles. In Section 3, we go into further detail about the QuartSet dataset. In Section 4, we demonstrate the usability of QuartSet by training an A2S transcription model with the real-instrument data alongside MIDI-synthesized data. Finally, we discuss potential usage for QuartSet and future work to expand upon QuartSet in Section 5. 2RelatedWork Due to the challenges of collecting real recordings, many transcription and source separation works for classical music choose to synthesize data instead. Most commonly, musical compositions that are available in MIDI format can be rendered as audio, using sound fonts and specifying note intensities and durations to make the renderings sound more realistic. Though despite its widespread use and easy application, MIDI as a format is limited in how much information it can convey. Information like expression markings, instrument information, and note groupings that would be important to a musician cannot be directly encoded in MIDI, and thus would not a"ect the audio rendering. Although synthesizing audio can create a large enough dataset for deep learning-based MIR tasks, it also sacrifices subtle aspects that contribute to the musicality. Nuances like tempo fluctuations, imprecise intonation, and varying interpretations of dynamics are all likely to occur with musicians of any caliber. However, audio synthesized from MIDI is quantized in a way that omits such imperfections. If a model is trained only on synthesized data, it may not be able to properly handle such nuances. MIDI is also a common data format for the ground-truth scores used in A2S transcription. As previously mentioned, MIDI data is limited to the basic information for pitches that occur throughout the composition. So notations that Proc. of the 17th International Symposium on CMMR, London, UK, Nov. 3-7, 2025 583 QuartSet: A String Quartet Dataset of Real Instrument Recordings are typically shown on a musical score, (e.g., dynamics, articulation, techniques) are not able to be rendered from MIDI data. Another common format for music transcription is MusicXML, which has the capability to encode the expressions that MIDI does not. Despite this advantage, MusicXML is more challenging to generate directly in an A2S context. Approaches like that of Nakamura, et al. [11] will instead train their transcription model to output a piano-roll representation of the music that can be saved as a MIDI file. This file is then opened with a score typesetting software, such as MuseScore or Sibelius, which can be used to, finally, export the score as a MusicXML file. 2.1 Datasets EnsembleSet [16] is an audio collection of classical compositions for the task of small ensemble source separation. The audio was rendered from both MIDI and MusicXML data that were publicly available. Using the BBC Symphony Orchestra sample library (BBCSO), di"erent articulation were rendered parsed form the MusicXML data. These articulations were encoded as MIDI keyswitches in the -1 octave. Overall, EnsembleSet was limited by the amount of available MusicXML files. Of the 80 compositions they sourced, only 30 were rendered from MusicXML. Meaning, these were the only pieces that had complete articulation data. For the remaining pieces, the music notation software that was used to import MIDI files would automatically select either staccato or legato articulation based on the individual note lengths. SynthSOD [5] is an audio dataset specifically designed for the source separation of full orchestras. Its contents were synthesized from MIDI scores sourced from the Symbolic Orchestral Database (SOD)1using the BBCSO library. Because of the aforementioned lack of expressive information when using MIDI, the authors created a pipeline that added those aspects into the audio rendering. To simulate di"erent dynamics, the common dynamic markings were randomly selected from certain values for MIDI velocity. So, the quietest dynamic (ppp) would have a velocity in the range of [1, 16), while the loudest (!f)wouldhave the range of [112, 127]. Tempo keywords, such as largo,andante,andpresto, were similarly given certain ranges for beats per minute (BPM) and randomized from those ranges. Articulations for individual notes were also randomized in acontrolledmanner.Eachinstrumentwasassigneddi"erentprobabilitiesfor each type of articulation. For example, a violin note had a 0.1% chance to be pizzicato, whereas a cello note had a 5% chance. The ranges and probabilities for these random augmentations were made with advice from orchestral music experts to ensure realistic outcomes. Kim, et al. [8] acknowledged the lack of available data for the source separation of string quartets and proposed a dataset rendered from MusicXML. It includes eight complete string quartet scores ranging from the Classical to Modern eras. They replicated the articulation markings from the score in the audio using a fixed process. Notes marked as staccato had their durations halved, 1https://qsdfo.github.io/LOP/database Proc. of the 17th International Symposium on CMMR, London, UK, Nov. 3-7, 2025 584 E. Rumbold and G. Tzanetakis and quartered for staccatissimo. The transition between notes that are slurred together is smoothed out by interpolating the amplitude and noise parameters, as shown in Figure 1. Dynamic notations are interpreted linearly and stored in MIDI as expression controls. Continuous dynamics, such as crescendo,are handled by estimating the duration and changes based on the starting points. Fig. 1. Example of linear interpolation to represent a slur. Bach10 [4] is a collection of ten four-part chorales written by J.S. Bach, each being about 30 seconds long. The audio is of real performances by a violin, clarinet, tenor saxophone, and bassoon. Each musician was recorded in isolation while wearing headphones to hear the others. The corresponding scores are comprised of publicly available MIDI files. The University of Rochester Multi-Model Music Performance Dataset (URMP) [9] is comprised of 44 small ensemble pieces assembled from coordinated, but separately recorded, performances of each part of the instrumentation. The most significant challenge faced in the creation of URMP was synchronization of the individually recorded performances. To address this complication, the authors recorded the audio and video of a conductor leading a pianist performing each piece. This resulted in a performance with natural variations in tempo and dynamics as indicated by the conductor. Then, the musicians would watch and listen to this conducted sample while they recorded their part of the piece. The corresponding score for each piece is provided in MIDI format along with videos of the assembled performances. Works on the separation and transcription of singers frequently use multiple datasets of real vocal ensemble recordings [3], [17], [19]. Commonly used vocal datasets include Bach Chorales2,BarbershopQuartet 3, Choral Singing Dataset [2], and Dagstuhl ChoirSet [15]. Each dataset is rather small, but they are often combined into one dataset large enough for deep learning. The use of this collective dataset also enhances the ability to generalize to unseen data due to the various genres, ensemble sizes, and recording environments represented. 2https://www.pgmusic.com/bachchorales.htm 3https://www.pgmusic.com/barbershopquartet.htm Proc. of the 17th International Symposium on CMMR, London, UK, Nov. 3-7, 2025 585 QuartSet: A String Quartet Dataset of Real Instrument Recordings 2.2 Motivation for QuartSet EnsembleSet [16] and the dataset presented by Kim, et al. [8] are both compilations of classical string quartets that were rendered from MIDI scores. They do their best to approximate realistic sounding audio by using the BBCSO sound library and defining fixed interpretations of articulation and dynamics, respectively. However, these methods of augmentation are not able to fully replicate the finer aspects of a real performance. For example, staccato notes played by a musician may vary slightly in duration or onset intensity. But when notes are augmented by a sound from the BBCSO library or specified as being half the duration of a legato note, there would be no variation at all. Of the datasets mentioned in Section 2.1, Bach10 [4] and URMP [9] are two that feature real instrumental performances. Bach10 is small, containing only ten recordings of about 30 seconds each. It also uses four instruments with very distinct timbres: violin, clarinet, tenor saxophone, and bassoon, which would not be optimal for the task of source separating closely related instruments, like a string quartet. The recordings in URMP have a variety of ensemble sizes and instrumentation combinations. Of the 44 arrangements, eleven feature only string instruments, and just five are string quartets—not nearly enough data for a deep learning training. Also, the scores provided with each of these datasets are in MIDI format. So, even though the recorded performances have musicality that synthesized audio would not, the scores do not reflect that. In creation of QuartSet, we prioritized the aspects of the music that are omitted from MIDI-based options. We did this by using real recordings of string quartet performances and by collecting scores in a format that is capable of representing the various dynamics, articulations, and embellishments written by the composer that are interpreted in the performance. 3QuartSet 3.1 MusicNet In this work, we introduce QuartSet4. This dataset is derived from MusicNet [18], a collection of freely-licensed classical music recordings. These recordings feature a variety of studio and microphone conditions, as well as a variety of ensemble groupings—from solo performances to octets of wind, string, and keyboard instruments. For this dataset, we focused on the string quartets. MusicNet contains the complete recordings for 56 string quartet pieces, amounting to 405 minutes of audio in total. These quartet recordings were sourced from the European Archive, Museopen, Musicians from Marlboro, and the Pascal String Quartet. An overview of the composers and musical keys for these pieces is shown in Table 1. For each recording, we manually selected ten excerpts, each 3-12 measures in length. 4https://github.com/erumbold/QuartSet Proc. of the 17th International Symposium on CMMR, London, UK, Nov. 3-7, 2025 586 E. Rumbold and G. Tzanetakis Table 1. Summary of composers and keys of string quartets from MusicNet. Composer Minutes Tracks Keys Tracks Keys Tracks Beethoven 214.2 28 F major 14 D major 3 Mozart 79.4 12 Eb major 8 Aminor 3 Dvo!ák 55.7 8 C major 6 E minor 3 Ravel 27.4 4 A major 5 Dminor 2 Haydn 14.8 3 Bb major 5 Fminor 2 Brahms 6.35 1 G major 4 C minor 1 3.2 Kern Notation Scores The scores of QuartSet are provided in Humdrum Kern notation due to its ability to easily represent polyphonic music. The plain-text style also allows for straightforward manipulation of the score via simple text replacement. In addition to representing the pitches and durations of the notes, Kern notation is also capable of showing markings for dynamics, articulation, and various ornaments. Fig. 2. A Kern-format score representing the adjacent Western notation music score. In a Kern text file, each column is associated with one source, and each line shows the events that occur and how they align with the other sources. Figure 2 shows an excerpt of a Western music notation sheet along with its corresponding Kern representation. Each column, or spine, is associated with one instrument. Spines labeled **kern dictate the note names, durations, and any articulation or Proc. of the 17th International Symposium on CMMR, London, UK, Nov. 3-7, 2025 587 QuartSet: A String Quartet Dataset of Real Instrument Recordings ornaments. Spines labeled **dynam show the dynamic markings for each source as they occur in relation to the notes. It is common for string instruments to play multiple notes at once. When the notes have the same duration, they will be written in Western notation with a shared note stem. If the simultaneous notes have di"erent timing, their stems will be separated. In Kern notation, the spine will be split, indicated by the starting token *ˆ,andtheendtoken*v. Notes played at once with the same duration are listed in the spine, separated by a space character. Examples of these notation di"erences are shown in Figure 3. Fig. 3. Examples of polyphony from one source instrument. The green section shows polyphony where the notes have di"erent durations. In Western notation, the note stems are split. In Kern notation, the split is preceded by *ˆ, then followed by *v when the parts merge. The blue section shows polyphony with shared note stems in the Western score and multiple notes within one spine in the Kern score. Several compositions from MusicNet have complete Kern scores readily available through the humdrum-data repository5.Wecopiedthesescoresandedited them to match the selected excerpts from the recordings. Some compositions that lacked an available Kern transcription had a score available in MusicXML format instead. We converted the MusicXML representations into Kern notation using the Verovio Humdrum Viewer6. The remaining pieces without any available scores were transcribed manually. 5https://github.com/humdrum-tools/humdrum-data 6https://verovio.humdrum.org/ Proc. of the 17th International Symposium on CMMR, London, UK, Nov. 3-7, 2025 588 E. Rumbold and G. Tzanetakis 3.3 Content Table 2. List of available metadata for QuartSet by category. Identity Score Audio Dynamics Articulation Ornaments ID Tempo Duration (s) cresc. decresc. Staccato Grace Note Composer BPM Source ppp pp Legato Trill Composition Measures pmpStaccatissimo Turn Movement Lowest Note mf f Marcato Mordent Catalog No. Highest Note ""fAccent Fermata Chords sf fp Dotted Legato Double Flat Clef Changes fz sfz arco Double Sharp rf pizz. Table 3. Examples from the metadata table, stored in quartset-metadata.csv. ID Composer Composition Movement Highest Cresc. 1788-0 Mozart No. 19 in C major 1. Adagio - Allegro Db6 Yes 1790-7 Mozart No. 19 in C major 3. Menuetto - Allegro E6 1919-4 Dvo!ák No. 10 in E-flat major 4. Finale Ab6 Yes 2104-1 Haydn No. 53 in D major 1. Allegro moderato G6 2315-8 Beethoven No. 15 in A minor 3. Andante A6 Yes The metadata for this dataset were partially taken from those provided with MusicNet. We reformatted this information and updated it to include metadata related to the Kern scores. All column labels are listed in Table 2, and a few examples from the metadata are shown in Table 3. The audio for each track is stored in a file named with the track ID. All tracks are mono .wav files with a sample rate of 44,100 Hz and a bit rate of 32 kbit/s. Each score is stored in a .krn file named with the corresponding track ID. 4 Experiment We tested the QuartSet dataset using an audio-to-score transcription model from Arroyo, et al. [1]. This model is a Convolutional Recurrent Neural Network (CRNN), in which a block of convolutional layers determines the features for the task. Then a group of recurrent layers models the temporal dependencies of those features and passes them to a fully-connected network for decoding. The CRNN model is trained using Connectionist Temporal Classification (CTC) loss [6], which allows the network to be trained on unsegmented sequential data. Proc. of the 17th International Symposium on CMMR, London, UK, Nov. 3-7, 2025 589 QuartSet: A String Quartet Dataset of Real Instrument Recordings 4.1 Data Preprocessing In addition to QuartSet, we used a set of string quartets first presented by Román, et al. [14] and utilized in the training of the CRNN+CTC model. This set includes Kern scores sourced from the humdrum-data repository for compositions by Beethoven, Haydn, and Mozart. Following the steps taken by Arroyo, et al., these scores were first cleaned by removing ornamental symbols including fermatas, beams, stems, slurs, elisions, editorial marks, rest positions, and grace notes. Notes and rests, clef, key, time signature, articulation marks, and dynamic expressions are kept. The pieces were then randomly split into 3-6 measure fragments and synthesized at a sample rate of 22,050 Hz using appropriate soundfonts for violins, viola, and cello. The audio is fed into the model as a log Short Time Fourier Transform (STFT) representation with log-spaced bins and log-scaled magnitude, using A4 as the reference pitch, 48 bins per octave, a 2048-sample Hamming window, and a512-samplehopsize. To prepare QuartSet, we cleaned the Kern scores using the method described above, and resampled the audio recordings to 22,050 Hz before generating the STFTs. Both QuartSet and the synthesized dataset were divided into train (70%), validation (15%), and test (15%) sets. These partitions were made by composition, rather than by fragment. That is, all excerpts of a piece would be in the same set to avoid possible biases. 4.2 Evaluation Metrics For evaluation, we observed the Character Error Rate (CER) and Word Error Rate (WER) of the transcribed symbols. These error rates indicate the number of elementary editing operations (insertions, deletions, or substitutions) at the character and word levels, respectively. In this case, a word would be the symbols that indicate a single note in Kern notation. For example, 8.bb-, which represents Bω5as a dotted-eighth note, would be considered one word for the WER calculation. 4.3 Model This CRNN scheme starts with two convolutional layers that apply 16 filters of size 3→3with a stride of 2 in the frequency axis, avoiding pooling layers. The output is split in half to reduce computational costs before feeding the features into the recurrent stack, which includes two Bidirectional Long ShortTerm Memory cells with 1024 hidden units each. Finally, the recurrent output is processed by a single fully-connected layer. The model was trained with a batch size of 4, Stochastic Gradient Descent (SGD) optimization with Nesterov momentum of 0.9. The learning rate schedule has an initial value of 3·10→4, and an annealing figure of 0.91. The model is trained for 50 epochs with the goal of minimizing the WER metric on the validation set. Proc. of the 17th International Symposium on CMMR, London, UK, Nov. 3-7, 2025 590