A Review of String Instrument Synthesis Methods for Use in Interactive Systems

2.1 Scope and review criteria

The focus of this review is on string instrument synthesis methods for use in interactive systems. We envision interactive musical systems to empower users to provide live input (e.g., by 3D gestures), to receive immediate auditory feedback, and to adjust their behavior in response to the feedback. Depending on the intended application—ranging from live performance to sound design—the requirements for response time (latency) and implementation vary substantially. Latency should not be presented as isolated numbers but contextualized relative to perceptual thresholds—specifically, distinguishing between the 10‑ms ideal for imperceptible delay and the 30‑ms ceiling generally acceptable for playability (Jack et al., 2016; Wessel and Wright, 2002).

Based on the implementation status observed across the 67 included studies, we classify studies into three mutually exclusive tiers:

• Tier 1 covers fully implemented interactive systems where users directly control the synthesis in a closed loop.

• Tier 2 includes synthesis methods that reported real‑time implementation for interactive use but have not yet been integrated into an actual interactive setup.

• Tier 3 refers to offline systems and implementations whose reported latency or processing constraints prevent interactive use.

We considered studies in all three categories to capture the full breadth of technical methods.

To define the instrument scope, we categorize the acoustic instruments by their excitation mechanism: plucked (e.g., guitar, guqin), bowed (e.g., violin, cello), hammered (e.g., piano, dulcimer), and ‘other’ (e.g., the wind‑driven aeolian harp). We explicitly include the piano since its simulation fundamentally concerns string vibration and hammer–string interaction, which is highly relevant to string synthesis research.

2.2 Literature search strategy

We conducted a structured search across five major databases: DBLP, the ACM Digital Library, IEEE Xplore, SpringerLink, and Scopus. The search combined descriptors for instruments (e.g., ‘string instrument,’ ‘violin’), synthesis methods (e.g., ‘physical modeling,’ ‘neural audio synthesis’), and interactivity (e.g., ‘real‑time’, ‘HCI’). We used Boolean operators to combine these sets. We restricted the search to English‑language publications from 1983 to 2025, as 1983 marks the introduction of the Karplus–Strong algorithm (KSA) (Karplus and Strong, 1983), which laid the foundations for modern digital string synthesis.

2.3 Inclusion and exclusion

We selected publications for inclusion if the proposed methods were explicitly designed to synthesize acoustic string instruments and were aligned with one of the three interactivity tiers described in Section 2.1.

Conversely, we applied specific exclusion criteria to keep our study focused. First, we excluded works dealing exclusively with electric string instruments (e.g., electric guitar). Second, we omitted publications focusing solely on signal analysis or recognition without a synthesis component. Third, we excluded synthesis methods based purely on offline sample playback without generative capabilities, as well as studies providing insufficient methodological detail. Finally, regarding data‑driven methods, we excluded general NAS methods unless they provided explicit empirical evidence of training or evaluation on string instrument datasets. Peer‑reviewed journal and conference papers formed the primary sources of the review. We considered preprints only if they provided complete technical descriptions and evaluations of string instruments and if no peer‑reviewed version was available.

2.4 Screening process

We followed a two‑stage screening workflow. First, we removed duplicates and screened titles and abstracts regarding eligibility criteria. Second, we assessed the full texts. The database search retrieved 2,360 records (ACM: 540; IEEE: 410; SpringerLink: 630; Scopus: 780), and, through scanning their bibliographies, we identified an additional 54 records. After removing 1,098 duplicates, 1,316 records entered title and abstract screening, of which 1,096 were excluded. We assessed the remaining 220 full texts and excluded 150 for one of the following reasons: no string instrument focus (38), analysis only without synthesis (29), reliance on offline sample playback (41), electric string instruments only (18), insufficient methodological detail (12), not written in English or not peer‑reviewed (7), and full text unavailable (5). As a result, we obtained 67 studies for our scoping review. Specifically, Tier 1 comprised 24 studies, Tier 2 comprised three studies, and Tier 3 comprised 40 studies.¹

2.5 Data extraction and categorization

For each included publication, we extracted structured data focusing on synthesis methods and target string instruments. In terms of implementation, we recorded information on real‑time capability and interactivity integration status, which informed our three‑tier categorization. Additionally, we further catalogued qualitative and quantitative evidence regarding fidelity, responsiveness, controllability, and adaptability. We provide the complete database in Table S1 of the supplement to support the subsequent comparative analysis. Among the 67 reviewed studies, several employed more than one synthesis method. Consequently, we conducted analyses at the method level (n = 71), exceeding the total number of studies.

3.1 Abstract digital sound synthesis

ADSS methods generally lack a physical basis or real‑world model, relying instead on perceptual and mathematical principles (Bilbao, 2009). We classify the following ADSS methods based on their primary components, such as oscillators, filters, and data tables.

3.2 Physical modeling synthesis

PMS methods attempt to simulate the acoustic mechanisms underlying the sound production rather than directly modeling the sound signal itself (Uncini, 2022). The underlying physical behavior is defined by a coupled set of partial differential equations (PDEs), which need to be (numerically) solved to produce the output waveform. We classify the PMS methods by their approach to modeling the vibrating system.

3.3 Neural audio synthesis

NAS uses DL to create and manipulate audio by training neural networks on audio datasets (Engel et al., 2017; Kalchbrenner et al., 2018). Unlike parametric methods, DL enables models to learn complex audio patterns, generate realistic sounds, and explore novel timbres, though it often demands extensive computational resources and large amounts of data (Natsiou and O’Leary, 2021).

We classify the following NAS methods by their generation paradigm and also include differentiable DSP (DDSP) as a complementary category, which provides physics‑guided structure and can be combined with the abovementioned traditional parametric methods within a hybrid system.

3.4 Summary of synthesis methods

The overall synthesis methods not only differ in their modeling approaches but also in the supported excitation mechanisms, as shown in Figure 4. NAS and PMS each account for a substantial share of the reviewed methods (n = 44 and n = 39, respectively), while ADSS appears far less frequently (n = 12). The comparatively high NAS count partly reflects the composition of training datasets, which often pool instruments across categories or apply generic labels, such as ‘Strings’ in the NSynth dataset (Engel et al., 2017), rather than precise excitation mechanisms. While PMS methods—particularly DWS and MIS—dominate plucked instruments and DNS also appears in bowed instruments, NAS—led by DDSP and DM—is the dominant approach for bowed and hammered instruments.

Figure 4

Aligned with these patterns, the distribution of synthesized instruments is also imbalanced. Well‑known instruments such as violin and piano appear in multiple studies, while many traditional or culture‑specific instruments appear only once or twice. We provide full details by instrument, synthesis method, and interactivity tier in Table S1 in the supplement. In the following section, we examine the evaluation of synthesis methods across four dimensions: fidelity, responsiveness, controllability, and adaptability.

4.1 Definition of evaluation dimensions

For consistency, we stick to the abbreviations introduced by Castagné and Cadoz (2003): ‘J’ denotes criteria originally proposed by Jaffe (1995), while ‘PM’ denotes those related to physical modeling (PM) methods, with J1–J10 and PM1–PM10 each comprising 10 criteria. We assign each criterion to a single main dimension using a primary‑motivation rule, as shown in Figure 5. When a criterion could plausibly support more than one dimension, we place it where its primary goal or bottleneck lies in interactive use and treat effects on other dimensions as secondary without double counting. The four dimensions are as follows.

Figure 5

• Fidelity evaluates the resemblance of synthesized sounds to real instruments, including the ability to preserve source identity under variation (J5), to generate sounds comparable to real instruments (PM2), and to remain plausible under exploratory modeling (PM5).

• Responsiveness addresses the computational and temporal requirements for synthesis, including algorithmic efficiency (J6 and PM1), control stream bandwidth and transport constraints (J7), and the minimum unavoidable latency of the technique (J9).

• Controllability captures whether performers can shape pitch, dynamics, and timbre in a stable and musically meaningful way, considering intuitive mapping to musical attributes (J1), perceptible effect of parameter changes (J2), physical coherence of parameters where appropriate (J3), and well‑ behaved response to controls (J4), complemented by modular construction and incremental modeling (PM6) and by a mental model that supports anticipation and exploration (PM7).

• Adaptability refers to the ease of transfer across instrument categories and the transparency of internal structure, including the breadth of representable sound classes (J8), the availability of analysis or parameter‑derivation tools (J10 and PM9), the diversity of instrument categories and mechanisms (PM3), and the depth of structural modeling (PM8).

Following the primary‑motivation rule, we assign J7 to responsiveness, since its central aim is feasible online operation. We assign PM6 primarily to controllability, since it provides clear modules and composition rules for performance and incremental modeling. We treat PM7 as a key support for controllability. Moreover, we deliberately exclude PM4, which concerns generality beyond sound synthesis, including haptic and visual interactions. We also exclude PM10, which concerns the quality of the surrounding musician‑oriented environment, because these pertain to multisensory or ecosystem considerations rather than intrinsic audio synthesis performance along our four dimensions.

4.2 Fidelity

Fidelity is a key criterion in evaluating instrument sound synthesis. To evaluate fidelity, 41 studies report some form of subjective evaluation, 41 report objective evaluation, 25 employ both, and only 10 report neither. These highlight both the central role of fidelity and the substantial variation in the ways it is evaluated. This subsection reviews how such evidence is reported, without assessing the actual quality or outcome of the methods.

We identified both formal and informal subjective evaluation methods, as summarized in Figure 6. Formal methods appear in 22 studies. The most common evaluation involves rating tasks, such as Likert scale questionnaires (5 studies, e.g., Dong et al., 2018) and Mean Opinion Scores (MOS) (5 studies, e.g., Jonason et al., 2023). A more rigorous variation is the Multiple Stimuli with Hidden Reference and Anchor (MUSHRA) (4 studies, e.g., Kim et al., 2025). Additionally, discrimination tasks such as ABX listening tests (5 studies, e.g., Défossez et al., 2018) are used to evaluate fidelity directly. Despite these formal evaluation methods, informal methods remain nearly as prevalent (19 studies), most frequently in the form of informal listening conducted by the authors themselves (e.g., Florens, 2003). Crucially for musical instrument synthesis, where realism of synthesized sound is a primary criterion, only three studies (e.g., Liu, 2024) explicitly involve professional musicians, who provided expert judgments on realism, revealing a significant gap in expert‑based validation.

Figure 6

For objective evaluation, we identified over 30 distinct metrics reported in the 41 studies, reflecting a wide variety of methods to evaluate fidelity. Spectral analysis—encompassing spectral losses, mel‑spectrogram distances, and related frequency‑domain measures—collectively appears in 16 studies (e.g., Carrillo and Bonada, 2010; Erkut et al., 2001), making it the most frequently reported class of objective metrics. Among standardized single metrics, the most common is Fréchet Audio Distance (FAD) (11 studies, e.g., Baoueb et al., 2024; Hayes et al., 2021). This usage coincides with the rise of NAS, offering a way to measure the distance between feature embeddings to assess similarity without requiring waveform alignment. Its predecessor, Fréchet Inception Distance (FID) (2 studies, e.g., Zhang and Akama, 2024) evaluates spectrogram images rather than learned audio embeddings. The shift from FID to FAD signals a methodological adaptation to the specific nature of audio data. However, error‑based measures—including mean squared error (MSE) (5 studies, e.g., Yang et al., 2020) and root MSE (RMSE) (3 studies, e.g., Bitton et al., 2020)—collectively remain the widespread default (see Figure S1 in the supplement).

Despite these tendencies, cross‑paper comparisons remain tentative because evaluation methods are inconsistent. Both listening‑based evaluations and computational metrics are often used, but their designs and implementations vary substantially across studies. In many cases, subjective and objective assessments are not applied together, making results difficult to compare. These fragmented practices highlight fidelity as a widely recognized criterion yet also reveal a pressing gap in methodological standardization.

In addition to differences between evaluation methods, the observed fidelity trends are also influenced by the type of instrument being simulated. As discussed in Section 3, string instruments vary substantially in the complexity of their excitation and sound‑production mechanisms. As a result, fidelity evaluations reported in the literature reflect not only methodological choices but also the intrinsic difficulty of the target instrument. This provides additional context for the frequent association of PMS with high‑fidelity plucked string synthesis and for the increasing use of NAS methods in the simulation of bowed and hammered instruments, where capturing continuous excitation and subtle timbral variation poses greater challenges.

4.3 Responsiveness

Responsiveness denotes the ability of a synthesis method to sustain interactive play. Specifically, it requires producing artifact‑free audio under low delay while supporting a target polyphony, defined as the number of simultaneous notes the system must render. Delay arises both from audio input and output (I/O) buffering and from control‑to‑audio processing within the method. For this reason, any ‘real‑time’ claim is only interpretable when quantified in terms of latency (ms), buffer size and sample rate, host/driver, and hardware. Moreover, in NAS literature, real‑time performance is frequently quantified via Real‑Time Factor (RTF)—the ratio of computation time to audio duration. While an RTF of <1.0 confirms that generation speed exceeds playback speed, it does not guarantee low latency. A system can achieve excellent RTF values yet require large input buffers or block‑based processing, rendering it unsuitable for interactive performance. Without these specifications, reported metrics represent isolated data points that cannot be meaningfully compared across studies.

The literature provides little quantitative evidence of system responsiveness. Of the reviewed studies, only 20 report any timing‑related metrics, and none report both latency and buffer size. This fragmented reporting—limited to latency in six studies and CPU load in eight—prevents meaningful cross‑paper performance comparisons. When the application context is specified, it is typically limited to desktop or plugin environments, with mobile, web, or XR targets rarely mentioned. Given this lack of standardization, existing reports fall into three categories. The most direct are hardware reporting standards, including CPU load (Bank et al., 2010; Silvin Willemsen et al., 2019), polyphony counts (Florens, 2003; Passalenti et al., 2019), or measured control‑to‑audio latency (Bitton et al., 2020). A second group evaluates algorithmic efficiency, giving operations per sample (Bank et al., 2010) or comparing computational complexity to baselines (Shan et al., 2022). The third group infers suitability from implementation environment, treating synthesis platforms such as Max/MSP, Pure Data, C++/JUCE, or dedicated DSP hardware as evidence of interactive viability (Selfridge et al., 2024).

Despite the heterogeneity in quantitative reporting, clear trends emerge when categorizing systems by their implementation status. As shown in Table 1, PMS dominates the fully interactive landscape (Tier 1), accounting for the majority of real‑time performance systems (18 out of 31 PMS studies). This shows that physical models remain the main choice for low‑latency, playable virtual instruments in interactive systems. While NAS is strongly growing (32 studies), it heavily concentrates on Tier 3 (27 studies), highlighting a gap: despite high fidelity, most NAS methods currently operate as offline generators or high‑latency systems. Notably, among NAS methods, DDSP is one of the few methods in Tier 1. Tier 2 represents a smaller collection of algorithms that have achieved computational efficiency for real time but await interactive integration. The fact that more than half of the reviewed studies fall into Tier 3 highlights a substantial gap between advances in string instrument synthesis and their translation into interactive musical systems. While sound quality and modeling accuracy have advanced, few works engage with the constraints imposed by real‑time interaction.

4.4 Controllability

In contrast to fidelity, which primarily concerns the resemblance between synthesized and target sounds, controllability addresses the way in which users shape synthesis output. We examine controllability in terms of the parameters of different synthesis methods, emphasizing their type, number, and interpretability.

To structure this comparison, we group controllable parameters into four distinct categories. Signal‑level parameters capture directly perceptible musical dimensions, such as frequency (pitch), amplitude, filter coefficients, or envelope shapes, which are intuitive and widely adopted. Physical‑level parameters correspond to acoustically grounded variables—bow position or force, string tension, damping, and hammer hardness—thus providing interfaces close to instrumental performance. Latent‑level parameters describe model‑internal or structural variables without direct acoustic correlates, such as latent coordinates, or embedding dimensions, enabling powerful but less interpretable transformations. The remaining ‘other’ parameters resist consistent assignment and are often tied to task‑specific conditioning or implementation details.

Synthesis methods reveal clear differences with respect to this categorization (see Figure S2 in the supplement). ADSS methods provide signal‑level controls such as pitch, loudness, or filter settings. Their number is limited since they serve for intuitive manipulation and since there are few ADSS methods in our review. PMS methods primarily exhibit physical‑level variables, allowing for control of sound production in ways that support expressive, performance‑like control, though often at the cost of requiring domain expertise. NAS methods display the most heterogeneous profile. Signal‑level and other parameters are most prevalent, while latent‑level parameters appear exclusively in NAS, reflecting the less‑standardized nature of NAS.

Controllability thus depends both on the number of exposed parameters and on their cognitive accessibility. This accessibility is determined by the mapping strategy, where the ideal design maximizes expressivity while minimizing cognitive demand (Levitin et al., 2002). Previous research suggests a tri‑partite division in how current methods approach this goal: signal‑level parameters prioritize accessibility, physical‑level parameters focus on mechanical authenticity, and latent‑level parameters enable creative exploration. The choice of method thus dictates not just the sound quality but also the interaction style available to the users.

4.5 Adaptability

Adaptability captures whether a synthesis method can remain effective when moving from a specific instrument to related ones. ADSS methods manipulate signal‑level parameters to synthesize specific instrument timbres. Since such controls lack correspondence to physical structure, migrating to a new instrument (e.g., from violin to cello) necessitates reconfiguring the synthesis architecture rather than simple scaling (Liu, 2024). PMS methods rely on a shared physical structure that theoretically supports geometric scaling. However, adaptation is often constrained by the complexity of parameter mapping. PMS employs explicit parameters to precisely model the physical interaction between excitation and the resonator. While this tight coupling guarantees high fidelity for the specific target, it requires recalibration when migrating to close relatives (e.g., violin to viola or cello) as tension, damping, friction, and body responses change. Without retuning parameter mapping, systems may satisfy the pitch range of the target instrument yet drift timbrally or break at transitions (Florens, 2003).

NAS methods achieve adaptability through data‑driven generalization. Training on multi‑instrument datasets and conditioning on instrument labels or learned timbre tokens enables these models to learn a shared latent space across instrument categories (Maman et al., 2025, 2024; Villa‑Renteria et al., 2025). However, reliance on shared weights introduces inherent risks—most notably, identity bleed, in which distinct instruments become perceptually similar or converge toward an averaged timbre when the latent representation is insufficiently disentangled.

In summary, methods deal with adaptability in distinct ways, manifesting as architectural rigidity in ADSS, calibration complexity in PMS, and latent entanglement in NAS.

5.1 Cross‑dimension tradeoffs

Our four‑dimensional evaluation framework forms a coupled design space where advances in one dimension may introduce costs on another.

A central tension exists between fidelity and responsiveness. Higher fidelity typically arises from richer excitation and radiation models in PMS or from large models with long analysis frames in NAS. These configurations increase mapping complexity and runtime load, thereby compromising responsiveness. ADSS methods usually achieve high responsiveness by using simple components but often fall short in capturing transient behaviors of acoustic instruments.

Second, there is a tradeoff between controllability and adaptability. Methods relying on explicit parameters (both PMS and ADSS) achieve high controllability by mirroring the specific structural constraints of a single instrument. However, this limits their adaptability: since the synthesis architecture is tightly coupled to a specific synthesis mechanism, migrating to even a closely related instrument requires significant reengineering—whether recalibrating physical coefficients in PMS or redesigning signal characteristics in ADSS. NAS methods stand out by addressing adaptability through learned shared representations. While this allows a single model to span an entire category, it often comes at the cost of control transparency.

All these configurations are constrained by the available computing budget, as improvements along any dimension increase computational complexity, limiting real‑time interaction. From this perspective, the advances in computational efficiency are not merely engineering optimizations (Caillon and Esling, 2021; Richardson et al., 2023) but are essential for rebalancing these tradeoffs and figuring out feasible strategies for real‑time synthesis.

5.2 Method–instrument alignment

This section examines how different synthesis methods align with particular string instruments in sound synthesis tasks. The four introduced dimensions help determine the suitability of a method for a given instrument category, even though no synthesis method is inherently limited to a single category of instruments.

In practice, different excitation mechanisms shape the spectral and temporal characteristics of string instruments. These characteristics, in turn, interact meaningfully with synthesis strategies. From the 67 reviewed studies, we found co‑occurrence patterns between synthesis methods and instrument categories.

Plucked string instruments, such as guitars and harpsichords, produce short, impulsive excitations followed by a relatively stable modal decay. The resulting sound structure is often spectrally sparse and can be effectively modeled with linear resonant systems requiring relatively few parameters. PMS methods—particularly DWS and KSA—are well suited to this structure. These methods expose interpretable controls such as string tension, damping, and body resonance and support efficient, real‑time synthesis with stable polyphony. Furthermore, parameter mappings in such models can provide a useful starting point for closely related instruments through pitch‑range scaling, though full timbral adaptation typically requires targeted recalibration.

In contrast, bowed string instruments involve a continuous, nonlinear excitation resulting from the stick–slip interaction between bow and string. This process produces dense harmonic content and highly sensitive variations based on expressive gestures—such as bow speed, pressure, and position. While physical models are capable of reproducing these effects, they often require precise calibration and become difficult to control when extended to different instruments or playing techniques. NAS methods have shown promise in this context, particularly when trained on multi‑instrument datasets with structured conditioning (e.g., instrument labels or gesture encodings). In our review, although most NAS studies are limited to offline use (Section 4), bowed string synthesis is handled best by NAS methods that maintain interpretability of control signals and support real‑time rendering (Maman et al., 2025; Villa‑Renteria et al., 2025). These methods offer a scalable and adaptable solution for modeling expressive, gesture‑driven instruments. Hammered string instruments exhibit forceful, broadband excitations followed by complex, long‑lasting resonances. These characteristics demand accurate modeling of both transient detail and extended decay behavior. Hybrid synthesis strategies are particularly effective in this context: physical models typically represent the resonant structure, while additive or subtractive components capture percussive transients and transient colorization.

ADSS methods (including AS and SuS) also appear across all defined string instrument categories, typically not as standalone solutions but as complementary components. They are commonly used to shape specific spectral regions—such as high‑frequency partials or resonator coloration—that are difficult to capture with physical or data‑driven models. In many systems, these layers are essential to achieving perceptual realism, particularly when enhancing or correcting the output of other synthesis engines.

Importantly, the synthesis–instrument matchings described here are not restrictive. PMS can be extended to bowed or hammered instruments, and NAS can be effective for plucked strings—especially when control fidelity and real‑time requirements are properly addressed. Ultimately, the choice of synthesis method depends on how well its structure aligns with the instrument’s excitation characteristics and the broader constraints of the system, including latency, control affordances, and computational cost. The method–instrument matrix (as shown in Figure 4) derived from our literature review reinforces these patterns, showing consistent tendencies that can inform synthesis design choices under practical constraints. The following section addresses how these method choices interact with interactive system design factors.

5.3 Runtime constraints

The feasibility of interactive string synthesis depends not only on algorithmic design but also on runtime conditions. Across the reviewed literature, three deployment contexts stand out. In desktop and plug‑in environments, PMS methods (e.g., DWS, MS) are widely reported to achieve immediate responsiveness and scale effectively to polyphonic use. NAS methods are viable in this setting when inference is streamed in short chunks on accelerated backends.

AS and SuS methods remain feasible if partition sizes are carefully tuned. In contrast, XR, mobile, and embedded platforms face strict power and scheduling limits. Reports of stable playability in these contexts are scarce, and, when present, they rely on compressed or quantized NAS models, short analysis frames, or partitioned convolution. However, the scarcity of explicit timing information and platform documentation makes reproducibility difficult.

In sum, responsiveness is not a property of algorithms alone but of their interaction with deployment settings and hardware constraints. This observation also reflects a broader pattern: across the four dimensions, tradeoffs and inconsistencies emerge that limit comparability of studies. This points to fundamental challenges in how interactive string synthesis is evaluated and reported, which we summarize in the next section.

5.4 Challenges

Building on our analyses, we see four recurrent obstacles that jointly limit comparability, replication, and long‑term progress.

The first challenge concerns fidelity. Perceptual quality is often optimized using spectral or waveform‑based criteria, yet improvements in such metrics do not consistently translate into musically convincing results. Moreover, increasingly complex models can compromise real‑time playability, and user‑centered perceptual evaluations are reported inconsistently, making it difficult to assess when technical gains yield audible benefits for musicians (Worrall et al., 2024). The second challenge is the insufficient and inconsistent reporting of responsiveness (Section 4). This lack of transparency hinders meaningful comparison across implementations and obscures the tradeoffs involved in deploying synthesis methods on different platforms, including desktop, mobile, and XR environments. The third challenge relates to controllability. Control interfaces and parameter mappings vary widely across systems and are rarely released or documented in sufficient detail. In NAS, in particular, latent control variables are often hard to interpret, which limits transparent authoring, reuse, and systematic comparison between methods (Engel et al., 2020, 2017). Regarding adaptability, the literature exhibits a bias toward a small set of instruments, most notably violin and piano, with limited evaluation of cross‑instrument generalization. As a result, claims of robustness at the instrument‑category level may conflate true adaptability with instrument‑specific optimization.

5.5 Recommendations and outlook

We now propose guidelines across the four‑dimensional framework, addressing evaluation and reporting practices for fidelity and responsiveness, and design priorities for controllability and adaptability.

For fidelity, subjective evaluations should adopt standardized protocols with hidden references and anchors, and informal evaluations should move beyond author‑centric listening by explicitly inviting professional musicians to assess the quality of synthesized sound. For objective evaluations, normalizing scores against a fixed reference dataset would allow direct cross‑study comparison. For responsiveness, we propose the lightweight reporting card in Table 2 for standardized documentation of system performance.

For controllability and adaptability, parametric methods would benefit from abstraction layers or adaptive interfaces that translate parameters into musically meaningful controls, alongside automated or learning‑based parameter mapping to reduce recalibration when scaling across instrument families. Data‑driven methods require more semantically interpretable and musically grounded mappings for latent parameters, as well as more disentangled and instrument‑aware representations to support cross‑instrument transfer.

5.6 Limitations

This scoping review has several limitations. First, although the search strategy covered major databases and multiple keywords, it may have missed relevant studies, particularly unpublished reports, non‑English works, or research outside the scope of interactive systems. Second, unlike a systematic review, this study did not apply formal quality assessment or quantitative synthesis; our aim was to map the landscape rather than to rank methods. Third, the four‑dimensional evaluation framework reflects choices informed by the literature, but alternative frameworks may capture other aspects of interactive synthesis. Finally, the field of NAS is rapidly evolving, and findings—especially concerning real‑time deployment on mobile and XR platforms—are time‑sensitive. These limitations do not undermine the value of the review but highlight the need for updates and complementary studies.