Rhythm Guitar Tablature Continuation Through Picking Pattern Generation

1.1 Musical considerations

In this paper, we focus on rhythm guitar, that is, accompaniment guitar parts in WPM, commonly opposed to lead guitar, which plays melodies and solos (Nemeroff, 2024). Although rhythm guitar is often identified on entire tracks, it is better formalized as a function that can be recognized at the bar level, as suggested in Régnier et al. (2021). We adopt the same assumption in this paper.

Guitar tablatures primarily transcribe gestural information, unlike standard notation that conveys harmonic and melodic information explicitly. Indeed, tablatures directly convey string–fret positions as well as playing techniques like slides or bends. They also include rhythm information, either directly on the tablature staff or in a separate standard notation staff.

Guitar chord diagrams can also be provided along with the tablature, as shown above the staves in Figure 1. A chord diagram is a graphical representation of the hand position and fingering required to play a chord on the fretboard. Diagrams show which strings are played, as some can be muted (x symbol). The frets that need to be pressed are denoted by black circles or lines (in the case of barred chords) or an o symbol for open strings (no fret pressed on that string). Diagrams can also be represented by their text‑equivalent—for example, x.x.0.2.3.2 for an open D Major chord.

1.2 Outline and contributions

We begin this paper with a presentation of existing work in accompaniment tracks and guitar tablature generation (Section 2). We then describe our approach for rhythm guitar tablature continuation in Section 3 and introduce the concept of picking patterns and symbolic texture metrics for tablatures. Section 4 presents the two models we implemented for picking pattern generation, given a chord progression with chord diagrams: a rule‑based algorithm and a Transformer neural network. Both models are used on the DadaGP dataset, which is presented in Section 5, along with our pre‑processing steps. The models are evaluated and compared with tablatures from the dataset through an automatic quantitative evaluation and an online user study—where 54 participants evaluated three different continuations of five random samples—described in Section 6. Results are reported in Section 7 and discussed in Section 8.

The contributions of this paper are as follows:

1. a representation of guitar tablatures through what we have dubbed picking patterns that permits a new modular approach for rhythm guitar tablature continuation;

2. a rule‑based and a data‑driven model, both based on picking patterns for generating respectively consistent and deterministic or more varied and creative continuations. Both can conduct inference without a GPU;

3. a novel set of evaluation metrics designed for guitar tablature, based on symbolic texture and chord diagrams; and

4. a systematic evaluation of rhythm guitar tablature continuations, through both automatic metrics and a subjective user study.

All code for reproducing this paper’s experiments and results is shared publicly under a GPLv3 license at https://github.com/adhooge/Rhythm-Guitar-Continuation.

2.1 Accompaniment generation and conditional symbolic generation

Generating rhythm guitar tablatures is a form of accompaniment generation. Generating accompaniment has been addressed directly in the audio domain for generating full bass tracks (Pasini et al., 2024) or any track that is not a lead melodic line (Nistal et al., 2024a, b). In Parker et al. (2024), a stem is generated and integrated into a context mix provided as input, while Wu et al. (2026) showed they can generate multiple high‑quality audio stems from a musical or text prompt. Accompaniment generation can also focus on specific instruments, such as drums. Lattner and Grachten (2019) propose a method to generate kick drum tracks in the audio domain, conditioned on snare and bass tracks. Accompaniment tracks can also be generated directly in the symbolic domain, such as through MIDI files. In Makris et al. (2022), the authors propose generating a MIDI drum track, using all the other instruments as input. Dahale et al. (2022) also tackle drum accompaniment generation, but with the additional step of generating improvisatory fills. Multiple tracks can also be generated simultaneously, like in Ren et al. (2020) and Huang et al. (2019), where multiple accompaniment MIDI tracks are generated based on a singing melody. Similarly, Dong et al. (2018) took one track as input to generate four others of typical rock‑music instruments. Another approach is proposed in Ens and Pasquier (2020), where some tracks can be provided to generate the missing others.

Additional control possibilities can also be designed to provide agency to composers. In Gao et al. (2024), for example, a chord progression is provided along with the desired lyrics to allow for more structural and harmonic control on the generated song. Likewise, Choi et al. (2021) generate vocal melodies conditioned by a chord progression to ensure that the singing matches the harmonic context. Another common approach to condition generation is to use a prompt as input to a generative model. In Chen et al. (2024), the prompt measure contains its musical content along with high‑level information such as tempo or style. Other approaches like inpainting or overpainting also assist composers by suggesting alternatives to a preexisting musical excerpt (Hadjeres and Crestel, 2021; Ens and Pasquier, 2020; Row et al., 2023). Further control possibilities can be implemented to improve models’ usability, like in Chen et al. (2020a), that support user‑defined ‘pitch and rhythm snippets,’ or like in Malandro (2024), that implement a range of user controls such as horizontal (temporal) and vertical (chords) note density.

2.2 Guitar tablature generation

Papers that study automatic symbolic music generation—for instance, in MIDI format—usually consider various instruments, including guitar (Zhu et al., 2018). Such research can, however, overlook some peculiarities of guitar playing by not considering tablature notation or by not modeling instruments individually. Conversely, there has been previous work which focused specifically on guitar music and generating tablatures. In Dahia et al. (2004), for instance, the authors generate bossa‑nova guitar rhythmic patterns based on a chord sequence and a melody. Likewise, McVicar et al. (2014) generate full rhythm guitar tablatures conditioned on a chord sequence using an $n$‑gram model. Another constrained approach to tablature generation can be found in Sakai et al. (2024), where chords and melody are already written and a fitting tablature is found by minimizing the transition cost between string–fret pairs. These previous references use optimization‑based framings, but deep learning approaches have also been considered for guitar tablature generation. In Chen et al. (2020b), a Transformer model is trained to generate ‘fingerstyle’ guitar tablatures. For additional control, the authors define grooving patterns to guide unconditional generation toward a desired rhythmic style. In Sarmento et al. (2023a, b), the DadaGP tablature dataset (Sarmento et al., 2021) is used to train Transformer‑XL models (Huang and Yang, 2020) with various control strategies such as using control tokens (musical style, instrumentation) or priming sequences (from one note to a full bar) to nudge the generation in the right direction.

An important limitation of previous work in tablature generation is the lack of chord progression conditioning for rhythm guitar tracks, with McVicar et al. (2014) being the only exception. In our work, we implement various control possibilities to enable a creative use of two suggestion models by supporting an input chord sequence as well as textural information.

3.1 Picking pattern generation

We propose to model rhythm guitar tablature continuation as composed of two successive steps linked to the two hands of the guitarist: i) choosing fretting hand positions (represented by chord diagrams) for an input chord progression and ii) selecting what we dub a picking pattern, which contains information on which strings should be played at each instant by the picking/strumming hand. Figure 2 illustrates two picking patterns applied to two possible chord positions of the Am chord.

Figure 2

Complementing existing work on chord diagram suggestion (Wortman and Smith, 2021; D’Hooge et al., 2024; Vélez Vásquez et al., 2025), our method focuses on picking pattern generation and considers that diagrams are provided along with the chord progression. An overview of our continuation pipeline is represented in Figure 3. Given an input bar in tablature notation, a continuation of three tablature bars is generated. We focus on four‑bar sequences, as this temporal scale in WPM has been shown to contain both consistent musical characteristics and contrastive parts (Bimbot et al., 2016). Our approach could nevertheless be extended to longer segments in future work. Once the picking pattern continuation has been generated, it is converted into an actual tablature by injecting the chord position information back into the output to assign strings and frets on each note. To further control the picking pattern generation, we also define conditioning signals representing textural and structural information, as described in Sections 3.2 and 3.3.

Figure 3

From a musical point of view, modeling each hand of the guitarist separately could be compared to splitting pitch and rhythm information, which was previously applied to symbolic music generation (Chen et al., 2020a; Dai et al., 2021). Applying this principle to rhythm guitar tablature generation has several advantages. First, it makes the approach modular and allows for updating its parts independently, for example, fine‑tuning only the picking pattern generation on a new dataset. Second, it allows for accepting varying degrees of user input, so composers are free to let the model suggest chord diagrams or to provide their own while still benefiting from picking pattern generation. This approach contrasts with previous work, where tablatures are generated directly as a sequence of tokens (Sarmento et al., 2021; Chen et al., 2020b) or as string–fret pairs for each note (McVicar et al., 2014; Sakai et al., 2024). In those approaches, guitarists cannot control intermediate steps of the tablature production and are therefore constrained to either modify the results manually or to generate alternatives if they are not satisfied with the output.

3.2 Describing tablature texture

This section presents a set of features for describing tablature texture. Texture features are used both for conditioning the generation (Section 3.3) and for estimating the distance between generated and reference tablatures in the evaluation (Section 6.1).

Tablature texture applies the principle of symbolic texture, as defined in Giraud et al., (2014) and Couturier et al., (2023) for Western classical music, to tablature. We similarly define vertical and horizontal texture features that respectively contain information about the pitches used and the rhythmic density. We use two horizontal texture features:

1. $\overline{N_o}=$ average number of onsets per beat, and

2. ${\sigma }_{IOI}=$ standard deviation of inter onset interval.

In addition, we use three vertical texture features specific to tablature content (D’Hooge et al., 2024):

1. $\bar{T}=$ average thickness: the number of strings played averaged by onsets, weighting the notes by duration;

2. ${\sigma }_{N_s}=$ number of strings standard deviation; and

3. $\overline{C_s}=$ average string centroid, obtained from the mean of the played strings numbers of each non‑rest onset. The final value is normalized by the number of strings.

As an example, taking the rightmost picking pattern from Figure 2 and attributing a weight of 0.5 for each eighth note, we obtain the following values:

We finally define horizontal and vertical texture variations between two consecutive bars $m-1$ and $m$ as:

where $T_H^m$ and $T_V^m$ are respectively the sums of horizontal and vertical texture features for bar $m$.

3.3 Conditioning controls

To extend controllability over the generated continuations, we propose to condition the generation using signals related to texture, chord positions, and structure, in addition to the input prompt. Texture conditioning consists of texture variation values as defined in Section 3.2. Specifying a conditioning texture aims to drive the model toward generating measures that may be unexpected or different. Conditioning on chord position diagrams aims to inform the model on which strings are played for each chord so that the proposed continuations do not use unwanted strings. For example, a C major chord played as a barred chord on the eighth fret uses all six strings (8.6.6.7.8.8), whereas, when played as an open chord, the sixth string (lowest) should not be played (x.3.2.0.1.0). Finally, we specify the index of the prompt bar passed through a modulo‑2 or modulo‑4 operator to add structure information to the conditioning. This informs the model about the position of the prompt within four‑bar cycles so that it can generate contrastive measures when appropriate (Bimbot et al., 2016).

4.1 Rule‑Based model

The rule‑based approach aims at reproducing the way composers and transcribers tend to use copy‑pasting features in tablature notation software to facilitate their writing process (Bacot et al., 2026). An example of an accompaniment part with a uniform texture where this copy‑and‑paste approach could ease writing is illustrated in Figure 4. Our model consists of a set of rules that apply minimal transformations to a tablature region to adapt it to a new chord diagram. Given a certain musical excerpt as input, the rules return a musical excerpt of the same duration, with the same rhythm, but with the string values (1 to 6) updated to match a new chord diagram.

Figure 4

Let ${\mathcal{S}}_P$ and ${\mathcal{S}}_D$ be the lists of the strings played in the prompt tablature excerpt and in the new chord diagram, respectively. $|\mathcal{S}|$denotes the length of the list $\mathcal{S}$ and $s_D^i$ denotes the number of the $i$‑th string in ${\mathcal{S}}_D$ ($0<i\le 6$), using the following mapping of open strings in standard tuning: $(E,A,D,G,B,e)\to (6,5,4,3,2,1)$. For example, let’s consider a chord that uses the three lower strings like a G power chord, or G5 (3.5.5.x.x.x), ${\mathcal{S}}_{D(\mathtt{G5})}=[6,5,4]$ and:

Based on this representation, we define several rules to map the picking pattern of the prompt measure to the next chord diagram. The first rule is fairly simple: regardless of how many strings differ between ${\mathcal{S}}_P$ and ${\mathcal{S}}_D$, if the chord is strummed entirely in the original bar, then all strings of the new diagram are strummed in the copy. However, in many cases, strings of a chord may be strummed independently (e.g., arpeggios, or bass notes played separately). In those situations, we map the strings from ${\mathcal{S}}_P$ to ${\mathcal{S}}_D$ according to the following rules:

• If $|{\mathcal{S}}_P|=|{\mathcal{S}}_D|$, for instance when changing from C Major (x.3.2.0.1.x) to D Major (x.x.0.2.3.2), we apply the mapping $s_P^i\to s_D^i,i\in ⟦0,|{\mathcal{S}}_P|⟧$. Note that, if ${\mathcal{S}}_P={\mathcal{S}}_D$, the mapping is the identity function;

• if $|{\mathcal{S}}_P|<|{\mathcal{S}}_D|$, the lowest strings are mapped to one another $s_P^0\to s_D^0$, and all other strings are chosen based on their distance to the highest string: $s_P^i\to s_D^{|{\mathcal{S}}_D|-(|{\mathcal{S}}_P|-i)}$.

• if $|{\mathcal{S}}_P|>|{\mathcal{S}}_D|$, we map the lowest strings together: $s_P^0\to s_D^0$, and the other strings are mapped according to the following rule:

  $$s_P^i\to \{\begin{array}{ll}{s}_D^{|{\mathcal{S}}_D|-(|{\mathcal{S}}_P|-i)}\hfill & ,\forall i\text{ where }{s}_P^i<\frac{s_P^{|{\mathcal{S}}_P|}+s_P^0}{2}\hfill \\ s_D^{\text{i}}\hfill & ,\text{otherwise.}\hfill \end{array}$$

If any duplicates appear in the mapping process, they are discarded.

Overall, these rules are designed to prioritize bass notes because of their critical function in chords. Then, strings are mapped with respect to their distance to the highest string in their respective diagrams so that arpeggio movements are replicated as closely as possible. We chose not to consider any conditioning information other than the chord diagrams in the rule‑based model to keep the rules simple and not err too far from a ‘copy‑and‑paste’ behavior. Including rules to account for texture controls is, however, a promising area for future improvements.

Once the picking pattern has been copied, frets are added on all notes based on the chord diagrams used at the corresponding time.

4.2 Deep learning model

In addition to the rule‑based model, we implement a deep learning model that can generate continuations in a data‑driven fashion unrestricted by our set of manually defined rules. Through stochastic generation, this model is expected to provide more varied and creative suggestions.

4.2.2 Data representation

Picking Pattern Tokenization: Picking patterns are encoded using a tokenization scheme adapted from Chen et al. (2020b). Like standard sheet music, tokenizing a tablature requires encoding the duration of each note along with its pitch, except that the pitch can be described indirectly with a string/fret pair. However, in our work, fret information is held within the chord diagrams, and picking patterns only gather string information with their musical timing. For this reason, we slightly simplify the tokenization proposed by Chen et al. (2020b): each new note is described by three tokens, that is, its onset, its duration, and the string it is played on. We also slightly increase the temporal resolution of the original tokenization as well and settle for a resolution of 12 ticks per quarter note, that is, 48 ticks per bar in 4/4. A summary of the vocabulary is provided in Table 1, along with an example of a tokenized excerpt given in Figure 5.

Table 1

Tokens in the vocabulary. Note: token:n‑m means that the value for token can be any integer between n and m inclusive.

Type	Tokens	Voc. Size
Metadata	<start>, <end>, <bar>, <PAD>	4
Note	string:1‑6, rest	7
Onset	onset:0‑47	48
Duration	duration:1‑96	96
Total		155

Figure 5

Conditioning Vectors: The structural information is represented using two one‑hot vectors of sizes 2 and 4 for the mod‑2 and mod‑4 measure numbers, respectively. The texture controls are represented with six float values: $[\mathrm{\Delta }{T}_H^2,\mathrm{\Delta }{T}_H^3,\mathrm{\Delta }{T}_H^4,\mathrm{\Delta }{T}_V^2,\mathrm{\Delta }{T}_V^3,\mathrm{\Delta }{T}_V^4]$. These 12 combined coefficients are passed through a single linear layer to obtain the conditional embeddings. Regarding diagrams, we only need to inform the model about which strings are used since we generate picking patterns and ignore fret information. We use an eighth‑note temporal resolution and therefore encode the strings used in a many‑hot matrix of size $32\times 6$ for the four measures in a sequence (as we limit our approach to 4/4 bars). This matrix is passed into another linear layer to obtain the diagrams embeddings.

5.1 Dataset

We conducted our experiments using the DadaGP dataset (Sarmento et al., 2021), which contains over 26,000 files in Guitar Pro² format. While the Guitar Pro software is proprietary, its native file format is supported by most music notation software used by WPM guitarists, including Musescore³ and TuxGuitar.⁴ The files in DadaGP usually contain several instrument tracks but always include at least a guitar track written in tablature format. This is the largest dataset of the sort available for research at the time of writing this paper. The dataset covers a wide range of WPM songs, with a bias toward rock and metal music, and is described extensively in Sarmento et al. (2021) and Sarmento (2024). The rest of this section details the preprocessing steps we implemented to use the dataset with our models.

5.2 Picking patterns extraction

Rhythm Guitar Identification: The DadaGP dataset contains all sorts of guitar tracks, with some being transcriptions of melodies and solos that are out of scope for this project. We extracted rhythm guitar content with a dedicated algorithm described by Régnier et al. (2021). Among the 2.9M bars of guitar tablature of the DadaGP dataset, 2.3M bars are labeled as rhythm guitar by the algorithm with at least a 50% confidence.

Four‑Bar Sequences Preparation: Our models are designed to work on four‑bar patterns, generating three measures given the first one. The patterns are obtained in each file with a one‑bar step sliding window, to ensure that the deep learning model can generalize to any part of the four‑bar cycles we observed. In terms of exclusion criteria, we drop any pattern that contains over 75% of rest time, or whenever the prompt bar is entirely silent. We also only consider bars in 4/4 time signature and leave the extension to any meter to future work. With those constraints, we obtain 2.6M patterns, of which 1M are unique.

Finally, we isolate a subset of the data where chord diagrams are annotated to conduct our second training phase, as described in Section 4.2.3. We identify 31,596 patterns with two to eight chord diagrams usable for the second phase of training (train: 21,987, validation: 4710, test: 4899), which leaves 987,149 patterns ($\sim$97%, train: 888,417, validation: 98,732) for the first phase. The stark difference in sizes comes from the fact that most patterns do not explicitly mention chord diagrams, even though they usually have chord labels transcribed. Future work could presumably benefit from a heuristic that identifies chord diagrams from tablature excerpts.

6.1 Computational metrics

To evaluate the picking patterns generated by the deep learning and rule‑based models, we define three computational metrics that measure the distance to the expected reference from the dataset.

First, we compute the Out‑of‑Diagram (OoD) notes ratio, that is, the ratio of notes played on strings that do not belong to the current chord diagram—for example, the lowest string in an open A minor: x.0.2.2.1.0.

Second, we use an edit distance to assess how close the generated token sequences are to the reference ones. The edit distance depends on the size of the sequences before padding, so we normalize values by using the reference sequence as the target length. Note that normalized distances are not necessarily smaller than 1, as an entirely wrong prediction twice the size of the reference would yield a value of 2. Finally, the texture distance (see Section 3.2) indicates the difference between the conditioning texture variation and the generated content texture variation. The final value is obtained using a Manhattan distance.

An example of a generated picking pattern and the reference tablature, along with the corresponding metric values, is shown in Figure 6.

Figure 6

6.2 User study

We gather subjective feedback on the picking patterns generated by conducting a user study via an online survey. We randomly selected five prompts from the test set, only enforcing that they come from different songs and have different textures, and asked participants to rate three proposed continuations: the reference tablature from the dataset, and the picking patterns generated by the rule‑based and deep learning models, rendered as complete tablatures using the chord diagrams of the reference.⁵ We use the ground‑truth diagrams to render tablatures because picking patterns alone are not standard and might confuse participants. This is somewhat equivalent to rating the rhythm guitar continuation pipeline in an ideal context where the diagrams are ‘perfectly’ generated (even though the ground truth as well can be considered non‑optimal at times D’Hooge et al. (2024). All samples are rated on 1–7‑point Likert scales according to four criteria:

1. Playability: is the shown tablature playable on guitar?

2. Consistency: are the strumming and rhythm consistent throughout the excerpt?

3. Interest: is the musical content shown interesting? (subjective appreciation of the tablature).

4. Usability: would the proposed tablature be usable in a performance or composition context without editing it?

We ask several other questions to assess the representativeness of our population study and the musical expertise of the participants. These questions are provided in Appendix B. During the final analysis, we use these answers to automatically cluster participants based on their musical practice and proficiency. The survey also included optional free‑text answers to let participants justify their ratings.

We conducted an a priori statistical analysis to determine the required sample size using G*Power (Faul et al., 2007). We based our estimation on paired t tests (since the continuations are evaluated on the same test data), aiming for a significance level of $\alpha =0.05$, a power $(1-\beta )=0.95$ and a default effect size $dz=0.5$ considering a two‑tailed distribution. This analysis identified a minimum sample size of 54 participants.

The study was deployed on an institutional instance of LimeSurvey (Figure 11) and advertised on public guitar and music forums as well as to acquaintances of the authors. All answers were gathered anonymously and are shared on the public code repository. Participants were not paid for completing this survey. This study was approved by the institutional review board of Université de Lille (project reference no.: 2024‑823‑S132).

7.1 Computational evaluation

The distribution of the metric values obtained on the test set are reported in Figure 7. We use boxplots to display the results because outliers (around 6% of the test set) shift the mean and standard deviation values. The statistical significance of the differences between the distributions was assessed through Wilcoxon signed‑rank tests with a Bonferroni‑corrected significance threshold $\alpha =\frac{0.05}{28}$.

Figure 7

Out‑of‑Diagram Notes: On average, 4.1% of the notes in the dataset itself are OoD. The deep learning model generates more OoD notes, but conducting the second phase of training (DL* vs. Base DL) and adding our custom diagram loss (DL vs. DL*) both contribute to significantly reducing the ratio of OoD notes. The rule‑based model, however, cannot generate any OoD note by design, as detailed in Section 4.1.

Edit Distance: The full deep learning model performs best, with a slightly lower edit distance distribution than the model without diagram loss (DL*, $0.001<\mathit{p}<\alpha$), which is itself lower than the model before the second phase of training. The rule‑based model results are more spread out, with many high outliers. The difference between the rule‑based and ‘Base DL’ models is not significant ($\mathit{p}>\alpha$).

Texture Distance: The rule‑based model does not include texture information in its rules. It therefore produces a spread‑out distribution of texture distance values, with a high average of 2.063 (median of around 1), which is more than seven times larger than that of the deep learning model (0.273, with a similar median). For the deep learning models, only the contribution of the second training phase is significant (DL and DL* vs. Base DL).

Overall, these results show that the two‑phase training we designed is efficient, as is the custom diagram loss we adopted. These quantitative measures also suggest that the deep learning model outperforms the rule‑based model in terms of edit and texture distances. However, the results of the subjective evaluation presented in Section 7.2 mitigate these quantitative results.

Ablation Study: To better quantify the contributions of our conditioning controls to the deep learning model performance, we also trained models with ablated controls. The corresponding box plots are shown on the right of each subplot in Figure 7. All distributions are significantly different (p $<0.001$) from the deep learning model. We report the results of three ablated models: without any pre‑attention, without texture controls, and without the first measure number controls. For the reference deep learning architecture, refer to Appendix A. We observe that the texture controls have a large impact on texture distance, as well as on performance in general, as the edit distance increases notably, even higher than the rule‑based model. The number of the first bar used as structural control also has an effect, though it is less noticeable. In fact, the ablated models are significantly different on all metrics (p $<0.001$), with the exception of the OoD ratio between the ‘No Bar Num.’ and the ‘No Texture’ models (p $=0.12$).

Decoding Robustness: Finally, we conducted experiments on the deep learning model with different $k$ values for top‑$k$ sampling during decoding, as well as with several random seeds. The results never showed any statistically significant differences in Wilcoxon signed‑rank tests (p $>0.1$), showing that our model is robust to probabilistic decoding and further supports the validity of our results and conclusions.

7.2 Subjective evaluation

Based on our a priori statistical analysis, we sought 54 participants for our user study. Most participants were men ($n=\text{41}$); three were women, and 10 did not disclose their gender. Participants were evenly distributed by age, with about 10 people in each 10‑year bracket from 18 to over 65 years, with an average age of 46 years. Finally, almost all participants learned music in Europe or North America ($n=51$). Figure 8 shows the ratings’ distributions for each sample in terms of consistency, interest, playability, and usability. We also aggregate the ratings across samples and represent the cumulated answers for each model in Figure 9. When considering individual samples, differences between the answers’ distributions are not always statistically significant, but cumulated answers show a clearer ranking with the reference being preferred over the rule‑based model, which is itself preferred over the deep learning model on all criteria except for ‘Interest.’

Figure 8

Figure 9

To further analyze the participants’ answers, we implement a linear mixed‑effects model (Pinheiro and Bates, 2000). We observed that the measured personal characteristics (age, gender, country where they learned music, and music proficiency level) of the participants do not significantly impact the ratings given. The results for the most significant variables are reported in Table 2, taking the first sample and the reference tablature as the base value. The participant variance is 0.503, and the residual variance is 1.6 (standard deviation of $\sim \text{1.265}$). The residual variance indicates that a substantial part of the variability in ratings remains unexplained, possibly because of unmeasured factors or random noise. In terms of fixed effects variables, the rule‑based model leads to a 0.156 decrease in ratings, while the deep learning model leads to to a 0.402 decrease. The effect of which sample is rated is not always significant, as only the ratings of sample 3 (−0.361) and sample 5 ($+\text{0.143}$) differ significantly from those of sample 1.

8.1 Comparing deep learning and rule‑based models

The lower edit and texture distances of the deep learning model (Figure 7) suggest that the generated continuations are more similar to the reference data. While defining additional rules for the rule‑based model to include texture controls could reduce this performance gap, the results of the user study show that the rule‑based model was appreciated despite its distance to the reference. Indeed, although the rule‑based continuations were not considered particularly interesting, they were deemed more usable and playable than those of the deep learning model, possibly because of their higher consistency (Figure 9). Based on feedback from the participants, we hypothesize that, while the deep learning model generates picking patterns close to the reference data, the small differences are detrimental to the overall musical quality of its outputs. For instance, minor differences in duration tokens can significantly alter the groove of a picking pattern. String token divergences can, in turn, alter playability, especially when producing OoD notes. Such specific issues are not captured by our implementation of the edit distance, which suggests that this metric should be used with care, as it does not necessarily reflect the musical qualities of the generated continuations.

In any case, the higher edit and texture distances observed in the rule‑based continuations highlight the fact that the composition of rhythm guitar parts in WPM goes beyond a simple repetition of texture on successive chords. Overall, the rule‑based model shows promise in assisting notation due to its consistency and predictability. Combined with delta‑based texture controls, the deep‑learning model in turn seems promising in assisting composition thanks to its ability to model realistic divergences from the prompt.

8.2 Out‑Of‑Diagram notes

The reference dataset surprisingly includes 4.1% of OoD notes. Many of these notes correspond to melodic ornaments or additional notes that do not strictly belong to the annotated chord symbol but serve a musical purpose.

Guitarists can also include muted strings in tablatures when they play dead notes, either intentionally to add unpitched notes, or for convenience by muting unwanted strings and strumming all strings regardless. These notes, which are currently labeled as OoD in our approach, highlight the fact that diagrams sometimes tend to be notated by transcribers to reflect hand positions rather than precise combinations of notes. Finally, some OoD notes in the reference dataset result from mistaken diagram notations or missing diagram updates, forcing the previous diagram to be considered as longer than it should be.

To better take into account meaningful OoD notes, our two models will need to evolve in opposite directions. Relevant OoD notes could be produced by the rule‑based model through the addition of dedicated rules. The deep learning model could in turn produce less irrelevant OoD notes by improving the custom diagram loss, or by adding a final output layer specifically for removing unwanted OoD notes. However, these objectives require a better understanding of OoD notes, as it is not straightforward to identify which are relevant or erroneous. Future research could address this question, possibly by conceiving dedicated classification models. Future work could also explicitly include dead notes in the modeling and tokenization of picking patterns to acknowledge their important role in rhythm guitar tracks. These future improvements could have an important impact, as the OoD notes in the deep learning continuations were described as confusing by several participants, while the rule‑based continuations consistently seemed more playable to participants due to the absence of OoD notes, according to their comments.

8.3 Evaluation process

Finally, the different results from our two experiments illustrate that the quantitative and qualitative metrics do not measure the same musical qualities of our models, a common issue in creative musical tasks. This supports the need for appropriate methods to evaluate models designed to assist musicians in their practice. Besides, it is possible that the preference for the rule‑based model might be partly explained by the design of our user study. Indeed, participants had to rate five samples generated by our models in a controlled and non‑interactive fashion. Although this common design choice facilitates evaluation, it slightly differs from the expected use of the models. Future work should consider an interactive evaluation protocol where participants can use the models in their favorite notation software freely. By offering users multiple continuations based on texture controls, or by varying temperature and sampling techniques, this evaluation could better assess the user preferences. Besides, the scope of the present subjective study is limited, because participants only rated five samples (15 continuations in total). Even if our statistical tests currently show only limited contribution of the samples themselves to the final ratings (the model type is a better explanatory variable, as shown Table 2), these conclusions might change when including more test samples. Finally, future work should also assess how using automatically suggested diagrams affects the quality of the final tablatures, as unplayable or imperfect diagrams could alter the subjective appreciation of the continuations.