Instrument-to-Instrument Translation (ITI)

• Instrument-to-Instrument Translation (ITI) is a framework that converts information from one instrument's domain to another while preserving task-specific invariants such as semantics, calibration, or structural details.
• Methodologies include rule-based translators and deep-learning architectures (e.g., GANs and U-Nets) that separate instrument-specific syntax from invariant content to ensure accurate domain mapping.
• ITI has practical applications in psychometrics, music, solar physics, and surgical vision, where it preserves meaningful constructs and operational validity across heterogeneous domains.

Instrument-to-Instrument Translation (ITI) denotes a class of translation problems in which information produced, encoded, or measured by one instrument is converted into the domain of another while preserving a task-specific invariant. In the cited literature, that invariant may be the constructs of a questionnaire, the semantics of an instrument protocol, the score-level content of a musical performance, the radiometric and structural behavior of a scientific image, or the segmentation utility of surgical imagery (Kante et al., 2020). This suggests that ITI is defined less by a single modality than by a common requirement: the source representation is rewritten into a target instrument domain without losing the properties that make the source scientifically, musically, or operationally valid.

1. Terminological scope and common structure

The term appears in technically distinct settings. In psychometrics, ITI refers to translating a validated measurement instrument from one language to another while preserving constructs and item intent. In formal language processing, it refers to translation between the command or data languages of heterogeneous instruments through an intermediate representation. In music and audio, it denotes instrument timbre transfer or instrument-conditioned resynthesis that preserves musical content while changing timbre. In solar physics, it denotes image-domain translation between instruments so that older or different sensors can be mapped into a homogeneous target domain. In surgical vision, it appears as domain translation that transfers segmentation capability from a labelled instrument/domain to an unlabelled one (Petrila, 2022).

Context	Source to target	Preserved property
Questionnaire translation	English instrument to French instrument	Constructs, meaning, style, semantic correctness
Formal translators	Instrument $A$ language to instrument $B$ language	Measurement semantics via an intermediate alphabet
Musical timbre transfer	Source performance to target-instrument audio	Score-level content
Solar image translation	One solar instrument domain to another	Radiometric and structural consistency
Surgical segmentation transfer	Labelled source domain to unlabelled target domain	Instrument structure and segmentation utility

A recurring distinction in the literature is between literal conversion and equivalence-preserving translation. The questionnaire literature explicitly prioritizes conceptual rather than strictly lexical equivalence; solar-image ITI is described as more than rescaling intensities; AdaTT adapts expressive details rather than blindly copying them; and coSegGAN constrains generation so that instrument shape is preserved rather than visually restyled without regard to downstream use (Kim et al., 14 Jun 2026).

2. Formalization and translation architectures

The most explicit formal account is Petrila’s general translator formalism, which defines a translator as

$T = (S, D, I, \delta, \theta),$

where $S$ is the source alphabet, $D$ the destination alphabet, $I$ an internal or intermediate alphabet, and $\delta$ and $\theta$ are source and destination rules (Petrila, 2022). In that formulation, ITI for instruments can be written as

$$T_{A \to B} = (S_A, D_B, I_{AB}, \delta_A, \theta_B),$$

with $I_{AB}$ representing instrument-independent concepts such as measurement types, units, configuration parameters, and states. The same paper stresses that translation requires “at least two steps of processing, one focused on source and other on destination,” with source-side directives such as #if and destination-side directives such as @if, plus generalized “class” directives that unify parsing, semantic interpretation, and output generation.

A different but compatible formalization appears in solar ITI, where the central mapping is written as

$B$0

with $B$1 and $B$2 denoting image domains of different instruments (Jarolim et al., 2024). In paired settings, the generator is trained on co-temporal observations; in unpaired settings, the framework follows CycleGAN-style bidirectional mappings $B$3 and $B$4 with adversarial, cycle-consistency, and identity losses (Schirninger et al., 19 Sep 2025). In music, AdaTT casts timbre transfer as conditional generation

$B$5

where the controls are pitch and loudness contours extracted from source audio, and the text prompt specifies the target instrument (Kim et al., 14 Jun 2026).

Across these formulations, the intermediate representation may be explicit or implicit. Petrila’s $B$6 is an explicit alphabet. In solar ITI, the latent representation of a U-Net or GAN generator plays the mediating role. In AdaTT, the control representation fuses quantized $B$7 and RMS embeddings, then modulates them with target-adaptive scaling. This suggests a common architectural pattern: source-specific syntax or signal structure is separated from instrument-independent content, then re-encoded in a target-specific form.

3. Measurement instruments and cross-language equivalence

A psychometric instance of ITI is provided by B-TAMBiT, “Back-Translation with an Adjudicator with Mono and Bilingual Tests,” developed to translate a validated English questionnaire on social network sites and information privacy into French (Kante et al., 2020). The procedure does not use the term ITI explicitly, but it is described as preserving the instrument’s conceptual, semantic, and functional properties so that the English and French forms can be treated as equivalent instruments.

The workflow comprises preparation and concept clarification, forward translation, back-translation, comparison, monolingual expert review, adjudication, mono- and bilingual tests, and finalization. The key constructs supplied to translators are SNSs, Convenience of Maintaining Relationships, New Relation Building, Enjoyment, Privacy Concern, Information Sensitivity, and Self-disclosure. Two independent bilingual translators perform forward and back-translation; a professor of Sociology and Anthropology reviews the French version for linguistic correctness and research appropriateness; and a bilingual adjudicator in ICT4D resolves discrepancies and makes minor corrections. Testing uses three French-only participants and two bilingual participants.

The distinctive feature of B-TAMBiT is that equivalence is operationalized through multiple checks. Back-translation is used to verify the translation and allow “modification of constructs and words without changing the semantic of the constructs or words that have no clear equivalence.” The adjudicator handles cases where one English term has “three to four corresponding words” in French and where sentence structure must change to avoid grammatical errors while keeping original meaning. The monolingual test checks clarity and naturalness in the target language; the bilingual test compares how the English and French instruments function in practice.

The paper reports translation outcomes rather than full psychometric validation. It states that “no major differences were found between the two versions of the instrument” in the bilingual test and that the final French version preserves “meaning, style and [is] linguistically and semantically correct” (Kante et al., 2020). It also notes that reliability, factor analysis, and measurement invariance are not yet reported. A plausible implication is that, in this branch of ITI, translation quality is established first through semantic and functional equivalence, with later psychometric analysis serving as a second-stage validation.

4. Musical and audio ITI

In music, ITI denotes changing instrumental identity while retaining the musical content of a performance. AdaTT defines instrument timbre transfer as generating audio that sounds as if a target instrument is playing while preserving the score-level content—melody and rhythm—of a source performance (Kim et al., 14 Jun 2026). The paper distinguishes timbral fidelity, timbral naturalness, and score-level content preservation, and identifies a central failure mode: structural controls extracted from source audio may contain instrument-specific expressive details that conflict with the target timbral identity. Its example is violin-to-flute transfer, where pitch-dominant violin vibrato conflicts with flute vibrato that is often loudness-dominant.

AdaTT addresses this with a Stable Audio Open backbone, ControlNet-based structural conditioning, Control Scale Predictors, and Text-Guided Control Scale Predictors. Pitch $B$8 and loudness RMS are quantized, embedded, fused, and then adaptively scaled by frame according to the target instrument text embedding. The target-adaptive fusion is

$B$9

so the model can attenuate or amplify pitch and loudness controls according to target identity (Kim et al., 14 Jun 2026). On ControlNet-based comparisons, AdaTT reports CLAP $T = (S, D, I, \delta, \theta),$0, $T = (S, D, I, \delta, \theta),$1 $T = (S, D, I, \delta, \theta),$2, KAD $T = (S, D, I, \delta, \theta),$3, TIM $T = (S, D, I, \delta, \theta),$4, NAT $T = (S, D, I, \delta, \theta),$5, STR $T = (S, D, I, \delta, \theta),$6, and QUL $T = (S, D, I, \delta, \theta),$7, with the best subjective scores and the best KAD among ControlNet-based models.

A symbolic and codec-based route to ITI appears in TokenSynth, which is not framed explicitly as ITI but is described as directly aligned with its requirements (Kim et al., 13 Feb 2025). TokenSynth is a decoder-only transformer that generates DAC audio tokens from MIDI tokens and a CLAP timbre embedding. The model supports zero-shot instrument cloning from reference audio, text-to-instrument synthesis from CLAP text embeddings, and text-guided timbre manipulation through embedding interpolation. Its basic conditional model is

$T = (S, D, I, \delta, \theta),$8

with $T = (S, D, I, \delta, \theta),$9 the timbre embedding and $S$0 the MIDI sequence. This makes TokenSynth an ITI system at the symbolic level: when the performance is available as MIDI, the same musical content can be rendered with different timbral conditions.

A complementary line of work concerns timbre representation rather than generation. A contrastive AST-based framework learns a shared timbre space for both single-instrument and mixture inputs, using triplet and InfoNCE objectives, and reports 81.7\% top-1 and 95.7\% top-5 accuracies for three-instrument mixtures under the full triplet loss (Vaillant et al., 16 Sep 2025). Because positive pairs are different sounds from the same instrument and negatives come from other instruments, the learned space is deliberately shaped around timbral identity rather than pitch or mixture context. This suggests that ITI in music increasingly decomposes into three subproblems: content preservation, timbre representation, and generative resynthesis.

5. Solar physics: image-domain ITI, calibration, and restoration

Solar ITI is the most explicit image-domain use of the term. The 2024 framework defines ITI as a general deep-learning method that translates between image domains of different solar instruments so that historic or degraded observations can benefit directly from more recent instrumental advances (Jarolim et al., 2024). The demonstrated applications are solar full-disk observations with unprecedented spatial resolution, a homogeneous data series of 24 years of space-based observations of the solar EUV corona and magnetic field, real-time mitigation of atmospheric degradations in ground-based observations, a uniform series of ground-based H$S$1 observations starting from 1973, and magnetic field estimates from the solar far-side based on EUV imagery.

The framework uses convolutional encoder-decoder generators, optionally with GAN components, reconstruction losses, perceptual losses, adversarial losses, and total-variation regularization. Evaluation includes PSNR, SSIM, correlation coefficients, FID, magnetic-flux agreement, and downstream solar-physics tasks (Jarolim et al., 2024). The paper emphasizes paired training on co-temporal observations or physically constructed degradations, explicitly to reduce hallucination risk and preserve physically meaningful structures.

A later application extends the ITI framework to unpaired image translations between Solar Orbiter/EUI/FSI and SDO/AIA in the 174/171 Å and 304 Å channels (Schirninger et al., 19 Sep 2025). Here the mapping is more than intensity rescaling: the model learns the joint distribution of intensities and structures so that FSI images are calibrated and enhanced to look and behave as if recorded by AIA. The generators are U-Nets, the discriminators are multi-scale discriminators, and the training follows a CycleGAN-style $S$2 and $S$3 regime with identity mappings.

For Earth-aligned test data, the paper reports FID improvements from 9.9 to 3.5 in 171 Å and from 11.9 to 4.0 in 304 Å, comparing baseline calibration to ITI translation (Schirninger et al., 19 Sep 2025). For light curves, the 171 Å MAE decreases from 14.4 to 0.6, and the 304 Å MAE from 2.9 to 1.6. The same study states that ITI calibration is independent of Solar Orbiter’s heliocentric distance and longitudinal separation, although at distances greater than $S$4 AU small-scale artifacts may appear and the authors recommend such products only for global studies. In this branch of ITI, translation serves simultaneously as calibration, enhancement, and data harmonization.

6. Surgical vision and cross-domain segmentation transfer

In surgical image analysis, ITI appears as translation of segmentation capability from one instrument/domain to another. The coSegGAN framework addresses the case where labels exist in a source domain $S$5 but not in a target domain $S$6, and it jointly trains generators, discriminators, and a segmentation network so that the generator maps labelled source images into target-domain appearance while preserving structure (Kalia et al., 2021).

The model uses generators $S$7 and $S$8, PatchGAN discriminators, and a U-Net segmentation model $S$9. The segmentation loss is

$D$0

so the translated target-style image inherits the source label (Kalia et al., 2021). Two additional constraints are central. A shape-preservation loss ties the generator to segmentation quality, and a structural latent-space loss

$D$1

encourages shared structural representations across domains.

The reported experiments cover Endovis, UCL, and a human prostatectomy dataset. In the Endovis $D$2 Surgery case, coSegGAN reaches 93.7\% Dice on the labelled source and 92.8\% Dice on the unlabelled target, with $D$3Dice $D$4 (Kalia et al., 2021). Other domain shifts are harder, especially when UCL is the source, but coSegGAN consistently improves target-domain Dice relative to baselines such as RASnet+, Ternausnet+, and U-Net_FL+. The authors’ interpretation is that segmentation-guided generation prevents the domain translator from altering tool shapes or introducing clinically unacceptable artifacts. This is a functional form of ITI: what is translated is not merely appearance, but the operational validity of the segmentation model.

7. Misconceptions, limitations, and research directions

A persistent misconception is to treat ITI as direct style transfer, literal translation, or global normalization. The surveyed literature argues against each of these reductions. B-TAMBiT rejects word-for-word literalism in favor of construct preservation (Kante et al., 2020). Solar intercalibration states that translation means more than rescaling intensities (Schirninger et al., 19 Sep 2025). AdaTT shows that preserving source structure too rigidly can impair target timbral fidelity when expressive details are instrument-specific (Kim et al., 14 Jun 2026). coSegGAN shows that visually plausible domain translation can still fail if instrument geometry is altered or segmentation becomes unreliable (Kalia et al., 2021).

The main limitations are likewise domain-specific but structurally similar. B-TAMBiT uses a small pretest sample and does not report quantitative psychometric validation (Kante et al., 2020). Petrila’s formal translator assumes that source and destination languages can be described as formal systems and does not explicitly model real-time constraints (Petrila, 2022). AdaTT is currently monophonic, depends on accurate $D$5 and RMS extraction, and does not explicitly preserve room acoustics or spatial cues (Kim et al., 14 Jun 2026). TokenSynth requires MIDI and therefore needs a separate transcription stage for audio-to-audio ITI (Kim et al., 13 Feb 2025). Solar ITI can hallucinate small-scale detail when the source instrument is too blurred or when the application departs from the training distribution, such as flares or off-pointing observations (Jarolim et al., 2024). coSegGAN still degrades under extreme domain shifts and in dark or blood-covered regions (Kalia et al., 2021).

The stated research directions are correspondingly broad. The questionnaire branch points toward reliability, factor structure, and measurement invariance testing (Kante et al., 2020). The protocol branch suggests extensible rule-based translators grounded in shared intermediate representations (Petrila, 2022). Musical ITI points toward polyphonic extension, richer control signals, and more nuanced text descriptions (Kim et al., 14 Jun 2026). Solar ITI points toward more instruments and wavelengths, public release of calibrated datasets, integration with physical forward models, and uncertainty quantification (Schirninger et al., 19 Sep 2025). Taken together, these directions support a general interpretation of ITI as a methodology for preserving invariant content across heterogeneous instrument domains while allowing instrument-specific realization to change.