AI Music Glossary

An experiment that removes or changes one component to measure its contribution.

A controlled experiment that removes or changes one component to measure its contribution.

A discrimination test where an unknown item X is identified as matching reference A or B.

Fine information about timbre, texture, transients, noise, and high-frequency content.

A model that predicts representations carrying information about how the audio should sound.

A discrete codec token that preserves information used to reconstruct sound.

Recomputing selected forward activations during backpropagation instead of storing all of them.

Memory used by intermediate forward values retained for backpropagation.

An adaptive optimiser using moving estimates of gradients and squared gradients.

An Adam-style optimiser with weight decay separated from the gradient-based update.

A small trainable module inserted into or attached to a frozen pretrained network.

A policy deciding whether, when, and at what priority a request may enter scheduling.

A complete service that accepts user intent or source material, runs models, validates and stores outputs, and provides a controlled user experience.

Distortion that occurs when frequencies are represented above the limit allowed by the sample rate.

A similarity result localised to corresponding temporal or symbolic regions in two items.

The relationship between data streams that shows which parts occur at the same time.

GPU memory currently occupied by live framework tensors.

The strength or size of a wave or recorded signal.

A deliberately selected stimulus used to establish a lower or otherwise defined point on a rating scale.

A skew introduced by the knowledge, vocabulary, assumptions, or instructions used during annotation.

A recorded estimate of how certain an annotation is.

The extent to which several annotators assign the same or compatible labels.

A filter used during downsampling to reduce frequencies that cannot be represented at the new sample rate.

A difference between checkpoint and model structures that prevents direct parameter loading or reuse.

The degree to which examples or hours come from a small number of artists.

A model or capability developed and used under documented permission, scope, control, compensation, attribution, and withdrawal arrangements.

A list of short phrases describing musical properties in a clip.

A serving pattern where a request is accepted and queued while the result is produced and retrieved separately.

A mechanism that lets a model combine information from selected positions in its context.

A trial or interaction designed to test whether a participant follows the procedure.

A training procedure that independently removes selected conditioning attributes according to configured probabilities.

Credit and source information required or provided for a work or dataset.

Documentation identifying contributors, systems, sources, and transformations involved in producing an asset.

An unintended audible defect introduced by recording, processing, encoding, generation, or decoding.

A natural-language description of audible content in an audio clip.

One separately stored stream of audio samples.

The process of checking, standardising, filtering, and documenting audio before model training.

Guiding generation with features or tokens derived from a reference recording.

A numerical vector representing selected properties of an audio segment.

An encoder that turns audio into a learned representation such as an embedding.

The perceived technical quality and acoustic plausibility of an audio signal.

A compact representation designed to identify the same or near-identical audio across files.

A compact representation designed to identify the same recording or excerpt despite selected distortions.

A sequence of encoded audio data stored inside a media container.

A discrete ID selected from a learned codebook to represent part of a compressed audio signal.

A discrete symbol representing part of a learned compressed audio representation.

A signal embedded into audio so an associated detector can later detect or recover information.

A signal embedded into audio so an associated detector can later identify or recover information.

A comparison between audio and text vectors in a joint embedding space.

A declaration explaining which rights or permissions a contributor claims to control.

A model containing an encoder that compresses data and a decoder that reconstructs it.

A PyTorch debugging mode that adds information for tracing problematic backward operations.

A computational procedure assigning a value to a defined property of outputs, references, conditions, or system behaviour.

A system for calculating derivatives by recording and differentiating program operations.

Generating a sequence one step at a time, with each selected token becoming context for the next prediction.

Generating a sequence one step at a time, with each selected output becoming context for later predictions.

A skew created when easily available data is treated as representative of the wider world.

The application of reverse-mode differentiation and the chain rule through a computational graph.

The reverse computation that calculates gradients of the loss with respect to parameters.

A design where each participant receives part of the full test while conditions receive balanced coverage overall.

The degree to which an adapted model preserves capabilities of its original checkpoint.

A simple reference method used to judge whether a more complex method adds value.

A group of training examples processed together.

The number of examples processed together in one batch.

Information available from positions before and after a generated region.

The number of bits used to represent each audio sample, determining how many numerical levels are available.

The number of bits used or transmitted per second.

Evaluation where listeners do not know which model produced each sample.

An evaluation where listeners do not know which system or condition produced each sample.

An uncertainty interval estimated by repeatedly resampling the observed evaluation items.

The resource or operation that most strongly limits system performance or capacity.

An adapter that projects features into a smaller dimension and then back to the model dimension.

The degree to which material joins naturally across an edge between generated regions or sections.

Perceptual and musical smoothness where generated edits meet preserved material.

A statistical model estimating relative item strengths from pairwise outcomes.

Creating two or more continuations from the same prefix by selecting different next tokens.

A harmonic or melodic gesture that creates a sense of pause, closure, or continuation.

Known positive and negative comparison pairs used to choose or test a decision threshold.

A rollout sending a small portion of production traffic to a new version.

The rule determining how many outputs are produced and which enter evaluation.

The rule determining which generated candidate is presented for evaluation.

A standard internal label used to group spelling variants, translations, or related source labels.

The level of evidence that a capability works reliably for a defined task, user group, environment, and governance framework.

A natural-language description of the content of an audio clip.

The process of producing a natural-language description from audio, tags, metadata, or a combination of sources.

The degree to which a caption is supported by the audio or trusted metadata it describes.

A sequence of model stages where the output of one stage becomes input to the next.

A sequence of model stages where one stage's output becomes the condition for another.

Record information such as artist, title, album, release date, identifier, and label.

A loss of previously learned capabilities caused by training on new tasks or distributions.

Generation of the next output using available past and current context without requiring future input.

A restriction that prevents a model from attending to future token positions.

A sequence model that predicts from earlier information without using future tokens.

A Creative Commons public-domain dedication tool intended to let a rights holder waive copyright and related rights to the extent legally possible.

The number and arrangement of audio channels, such as mono or stereo.

A saved record of model and optional training state.

The degree to which saved parameter tensors match a target architecture and configuration.

A policy determining which checkpoints are preserved and for how long.

An ordered set of chord labels, often aligned with bars or time positions.

A representation grouping spectral energy by pitch class while normally ignoring octave.

A time-based representation that groups musical energy into pitch classes.

Consistency and smoothness across separately generated streaming segments.

An audible or structural mismatch at the boundary between separately generated regions.

Creating a long output as several shorter regions.

Similarity between audio and text representations produced by a CLAP-style joint embedding model.

A distribution where labelled classes contain substantially different numbers of examples.

A comparison based on predicted class distributions from an audio classifier.

A method that combines conditional and unconditional diffusion predictions to adjust condition strength.

A training procedure that removes all conditions for a selected proportion of examples so conditional and unconditional behaviour can be learned.

Distortion caused when a signal exceeds the level a system can represent.

A token from the more important or earlier codec quantisation levels, carrying a rough acoustic representation.

A process that creates a rough or compressed representation before progressively adding detail.

A learned collection of vectors that a quantizer can choose from.

A learned collection of vectors that can be selected and represented by discrete token IDs.

A system that encodes data into a compressed representation and decodes it back into a usable form.

The part of a neural codec that converts compressed representations back into a waveform.

Audio reconstructed from a compressed codec representation.

Additional delay caused by provisioning a worker or loading a model before it can process a request.

The duration and temporal region used as one unit in a similarity calculation.

Rights and permissions connected with the underlying music and lyrics.

Reducing the amount of data needed to represent information.

An unwanted sound introduced or made more noticeable by compression and reconstruction.

A graph of operations and dependencies connecting inputs, parameters, intermediate values, and outputs.

The numerical work performed by data preparation, model operations, gradient calculation, optimisation, and evaluation.

A piece of guidance used by a model during generation, such as text, melody, lyrics, genre, or artist.

An input supplied to guide generation.

A test that removes, replaces, or shuffles a condition to measure its effect.

The degree to which output follows requested text, melody, harmony, timing, reference, or structural controls.

A situation where two or more controls encourage incompatible outputs.

An assumption that some information is no longer needed once another conditioning variable is known.

The probability of an outcome given specified context or conditions.

Providing extra information that changes the probability distribution over possible outputs.

Providing information that guides a generative model toward associated outputs.

Information given to a generative model to guide the output it produces.

Information such as text, melody, labels, or audio used to guide generation.

A component that converts a control input into features used by a generative model.

The transformations connecting a control input to the model's internal prediction process.

Text given to a generative model to guide what it produces.

An uncontrolled factor that makes it difficult to identify the cause of an observed difference.

Evidence that a person agreed to a stated use under defined conditions.

A cryptographically bound C2PA manifest containing assertions about an asset and its provenance.

A cryptographically bound C2PA manifest containing assertions about an asset and its provenance.

The information available to the model when it makes a prediction.

A gradual movement of the generated sequence away from its earlier prompt, pattern, or intended direction.

The number of earlier tokens supplied to the model for a prediction or continuation.

The amount of sequence information a model can process together.

The part of a sequence and conditioning information available for a prediction.

The amount of sequence and conditioning information directly available for one prediction.

New sequence material generated after a supplied prompt or prefix.

Generating new material after an existing musical prefix.

Continuing a pretrained model's original or closely related objective on additional data.

A training approach that brings related examples closer in a representation space and separates unrelated examples.

An agreement describing what a contributor provides, controls, permits, and promises.

The practical ability to influence a selected property of generated music.

The degree to which generated music follows a supplied condition.

Information used to influence one or more properties of generated output.

A relevant factor held constant across experimental conditions.

A defined list of allowed labels used to make annotations more consistent.

A method for comparing the direction of two vectors.

A comparison based on the angle between numerical vectors.

Assignment of shared infrastructure costs to products, users, requests, or other defined groupings.

Distribution of stimulus positions or sequences across experimental conditions.

The extent to which a dataset includes the groups, conditions, and variations needed for an intended use.

A selected time range taken from a longer recording.

A loss function that measures how well predicted token scores match the correct target.

A loss comparing model logits with correct discrete targets.

A table showing counts or measurements for combinations of categories.

A failure in which information outside the intended training data influences training or makes evaluation easier.

Held-out information entering training or model-selection decisions and producing optimistic evaluation.

Use of held-out information during training or model selection in a way unavailable under intended deployment.

The chain connecting a processed item with the earlier items and activities from which it came.

A component that selects, batches, and supplies examples to the training loop.

Collecting only personal information necessary for the stated purpose.

Distributed training where workers process different data and combine gradient information.

Information about the entities, activities, and people or systems involved in producing and changing data.

Generation of output matching training material under a stated representation, metric, and threshold.

Generation of content reproducing an identifiable portion or representation of source material.

An organised collection of records used for training, validation, testing, or analysis.

The distribution of examples across the categories and characteristics relevant to an intended use.

A systematic skew in examples, labels, collection, or evaluation that influences learned behaviour.

Structured documentation describing a dataset's motivation, composition, collection, processing, uses, and limitations.

Documentation describing a dataset's motivation, composition, collection, processing, uses, maintenance, and limitations.

A structured record listing dataset items, identifiers, paths, metadata, splits, and provenance.

A diffusion sampling formulation that supports alternative and often shorter generation paths using a model trained with a standard diffusion objective.

A denoising diffusion probabilistic model based on a forward corruption process and learned reverse transitions.

A queue receiving work items that exceeded their successful-processing attempts.

A state where processes wait indefinitely and no worker can make progress.

A structured record of a failure, reproduction, evidence, root cause, fix, and verification.

The controlled process of locating, explaining, and correcting unexpected system behaviour.

A predefined rule connecting evaluation evidence to an action such as release, revision, limitation, or rejection.

The part of a codec that turns the compressed representation back into waveform audio.

Software that turns encoded audio into sample values that can be processed or played.

The part of a model that converts a latent or compressed representation back into an output.

Finding and removing or grouping records that contain the same or nearly the same source material.

Detection and removal, grouping, or reweighting of exact or near-duplicate examples.

Estimating how a corrupted representation should change toward a cleaner and more data-like state.

One update that moves the current noisy representation toward a cleaner state.

A mutable name pointing to an immutable model or pipeline version for a deployment purpose.

A tensor separated from the automatic-differentiation graph so earlier operations do not receive gradients through it.

A generative model that learns to transform noise or a corrupted representation into structured data through iterative denoising.

A generative model trained to reverse a gradual corruption process.

An index representing the current noise level in the diffusion process.

A digital asset or model intended to recreate the identifiable voice of a particular person.

One position or coordinate in a vector.

A method for representing data with fewer dimensions, often so high-dimensional embeddings can be visualised.

A representation built from a finite set of token or code values.

A checkpoint designed to save and restore state partitioned across distributed workers.

A data-parallel method where each process normally holds a model replica and synchronises gradients.

A linked record of operations performed as one request crosses services and processes.

The degree to which generated outputs occupy relevant regions of a target space.

A difference between the data represented by training, validation, testing, or deployment inputs.

A measure comparing statistical representations of generated and reference collections.

Meaningful variation across valid generated outputs.

Adjusting a model to data whose distribution differs from the original training domain.

A difference between the distributions of earlier training data and new data or use cases.

A test where participants and interacting administrators are prevented from knowing stimulus identities during collection.

Combining several audio channels into fewer channels.

A regularisation method that randomly sets selected activations to zero during training.

A provenance design combining cryptographic binding with discoverable soft bindings such as fingerprints or watermarks.

Providing desired output length or timing information to the model.

Combining compatible requests arriving near one another into a larger inference batch.

The difference between the quietest and loudest signal levels that a system can represent.

Ending training or selecting a checkpoint based on held-out performance before the final planned update.

The degree to which generated music can be modified or integrated into a creative workflow.

The total number of examples contributing to one optimiser update.

The total examples contributing to one optimiser update across microbatches and workers.

The earlier information that meaningfully influences a current prediction.

A numerical vector produced by a model to represent learned features of an item.

The statistical arrangement of a collection of vectors in an embedding space.

A model that maps inputs into a numerical representation space.

Searching for nearby reference items using vector representations.

The coordinate space in which embeddings are placed and compared.

A model or part of a model that turns input data into a learned representation.

The part of a codec that turns audio into a smaller learned representation.

The part of a model that converts input data into a compressed representation.

A linked record of source assets, model and dataset versions, generation conditions, edits, approvals, and release transformations.

The ability to organise the final region as a musical conclusion.

The pattern of increasing, decreasing, or changing musical intensity across a piece.

A recorded specification of software packages and versions required by a run.

One pass through the examples assigned to a training dataset or sampler.

A process where an unwanted generated choice changes later context and increases the chance of further unwanted choices.

A standardised collection of inputs, references, procedures, metrics, and reporting rules.

A problem where the test data or metric gives an incomplete or uneven picture of model performance.

One separately defined aspect of model or output performance.

A fixed collection of conditions designed to test a model's intended range and known challenges.

The complete rules governing generation, preprocessing, measurement, analysis, and reporting.

The precise judgement a participant is instructed to make.

A structured specification of evaluation dimensions, scales, anchors, weights, gates, and decision rules.

A discrete symbol representing a musical action such as a note onset, note release, velocity, or time shift.

A sequence whose tokens describe musical changes such as note starts, endings, time movement, and velocity.

The strength and relevance of evidence supporting a capability or deployment claim.

A record or file that is byte-for-byte identical to another.

Two dataset items identical under a chosen canonical representation or hash process.

Equality of digital files or byte sequences under a cryptographic hash.

A generated sequence identical to a reference under a selected representation and range.

The predefined rule identifying which region of each output enters the test.

A controlled procedure designed to test a specific hypothesis.

A structured specification of settings used to execute a model run.

Recording run configurations, metrics, artifacts, code, and relationships.

A participant selected for relevant experience, training, or demonstrated perceptual sensitivity.

The mismatch between training on correct prefixes and generation from prefixes containing model-selected outputs.

The number of times an example or related content contributes to training updates.

The proportion of generations that violate defined requirements.

An observable sign of a problem that may have several possible causes.

A replicated relationship missed by the selected test.

A screening result that fails to identify a relevant relationship.

A non-replicated pair incorrectly flagged as replication under the evaluation definition.

A screening result that incorrectly flags unrelated material as a match.

A musical property stated in a caption even though it is not supported by the audio or trusted metadata.

Using frozen pretrained representations as inputs to a newly trained head or downstream model.

Systematic distortion in feedback caused by response selection, interface design, candidate filtering, or missing failures.

Checking that a file can be read and contains the signal and technical information expected by a pipeline.

A token from later codec quantisation levels that refines local sound and reconstruction quality.

Additional training that adapts a pretrained model to a narrower dataset, task, or behaviour.

A test confirming that a tensor contains no NaN or infinite values.

Suitability of a system or output for its defined intended use.

An exact attention algorithm that reduces memory traffic through tiled computation.

A representation describing which sections or structural roles should occur and in which order.

A callback invoked around a module's forward computation for inspection or instrumentation.

The computation that transforms inputs into model predictions.

The predefined process that gradually adds noise to clean training data.

A metric comparing distributions of evaluated and reference audio embeddings.

A distance comparing means and covariance structures of evaluated and reference audio embeddings.

Preventing selected parameters from receiving training updates.

The number of wave cycles completed each second, measured in hertz.

Updating all or nearly all trainable parameters in a pretrained model.

A distributed method that shards parameters, gradients, and optimiser states across workers.

A component that combines information from a new condition with the internal representation of a base model.

A change applied to the level of an audio signal.

Random numerical values drawn from a normal distribution.

The ability to perform usefully on relevant examples not used for parameter updates.

The difference between training performance and relevant held-out performance.

The degree to which generated music satisfies defined technical, musical, conditional, creative, operational, and user requirements.

A durable record representing one accepted unit of generation work.

The repeated process of predicting, selecting, appending, and predicting again.

A structured record of the model, inputs, settings, assets, transformations, and approvals associated with an output.

A structured record of the inputs, model, settings, transformations, evaluations, outputs, and approvals for one generation.

Connecting genre labels from one vocabulary or taxonomy with labels in another.

Consistency, development, and meaningful organisation across the complete structure of a piece.

A condition intended to guide the generated output as a whole.

Organisation and development of musical ideas across a complete piece.

A running count of optimiser updates.

A fixed and inspected batch used as a reference across runs or code changes.

A processor commonly used to accelerate parallel numerical operations in neural networks.

Device memory used to store tensors and state needed by GPU operations.

Information showing how model parameters should change to reduce the loss.

A collection of derivatives showing how the loss changes locally with model parameters.

Adding gradients from several batches before one optimiser update.

Adding gradients from several microbatches before one optimiser update.

Limiting gradient magnitude before an optimiser update.

Limiting gradient magnitude before an optimiser update.

A callback invoked when a tensor gradient is calculated.

Memory used to store derivatives for trainable parameters.

A measurement of the overall magnitude of a collection of gradients.

A measurement summarising the magnitude of a collection of gradients.

Scaling the loss and gradients to reduce lower-precision numerical underflow.

Selecting the highest-probability token at every generation step.

The reference answer or observation treated as correct for a training or evaluation step.

A correct earlier sequence taken from the training data.

Reporting model performance separately for relevant groups rather than only as one overall score.

A split keeping all examples from the same defined group in one partition.

A generation setting controlling how strongly classifier-free guidance emphasises the condition.

A setting controlling how strongly a conditional prediction influences generation.

A frequency component related to a sound's fundamental frequency that helps shape its tone.

Guiding generation with chord or tonal information.

A category with many training examples.

The unit for cycles per second, abbreviated Hz.

A reference-quality stimulus presented anonymously among evaluated conditions.

Generation divided into levels that operate at different representations, resolutions, or time scales.

Generation where one representation or model stage provides guidance for a later stage at another level of detail.

Generating or specifying high-level structure before creating lower-level musical events or audio.

An arrangement of representations or models operating at different levels of compression, detail, or time scale.

A structured study in which participants listen to audio and provide defined judgements.

A specific prediction about what an experiment will show.

An identifier used to prevent repeated submissions from creating duplicate effects.

A gradual loss of the musical characteristics established earlier in a generation.

The planned process for detecting, containing, investigating, communicating, and correcting a harmful or unreliable event.

Using a trained model to produce outputs without updating its learned parameters.

The time required for a trained model to produce an output.

A PyTorch execution mode disabling gradient-related tracking and selected autograd overhead.

Using a pretrained model without updating any parameters.

Voluntary agreement to participate after receiving understandable information about the study.

An asset or data item recorded as a component of, parent of, or input to another asset.

The random starting representation used during standard diffusion generation.

Generating or replacing a selected region while preserving surrounding material.

The information given to a model so it can make a prediction.

A combination of two or more characteristics analysed together.

Variation among outputs generated from the same condition.

Improving a partial or uncertain representation over several repeated steps.

Improving a representation through several repeated updates.

A durable waiting structure separating accepted work from available processing capacity.

A named stage describing the current lifecycle position of a generation job.

A shared representation space that allows different data types, such as audio and text, to be compared.

A directional measure of how one probability distribution differs from another.

A category or value attached to an example and used as a target or description.

A systematic problem in how examples are named, categorised, or annotated.

The person, system, catalogue, or model that produced an annotation.

A model that learns to predict sequences of tokens. In audio generation, the tokens can represent compressed sound rather than text.

Diffusion performed in a compressed autoencoder representation rather than the original data space.

A learned internal representation that stores useful information in a more compact form than the original input.

A learned compressed numerical space produced by an encoder.

A policy workflow using multiple checks at different stages of the generation pipeline.

A hyperparameter controlling the scale of optimiser updates.

A rule that changes the learning rate during training.

A set of terms describing how data or content may be accessed, copied, modified, or used.

Permission that allows specified uses under stated conditions.

Whether the permissions and conditions attached to different assets or datasets can operate together for the intended use and distribution.

A requirement or limit attached to a licence.

A person or organisation receiving permission under a licence.

A person or organisation that grants permission under a licence.

A capability made available to a restricted set of users, contexts, regions, or partners.

Practice and instruction used to familiarise participants with the task and scale.

Variation among listeners' ratings or preferences for the same material.

Variation among participants' judgements of the same stimulus or comparison.

Reduced attention or changed judgement caused by prolonged listening or repeated evaluation.

A predefined participant subgroup analysed separately because its background may affect judgement.

Consistency and plausibility across a relatively short span of music.

A condition whose value or influence changes across time.

Consistency among nearby rhythmic, melodic, harmonic, and timbral events.

Adding detail using nearby context rather than remodelling the complete long sequence.

An unnormalised score produced for one possible output.

Generation over a duration requiring persistent structure, development, transitions, and resolution.

A relationship between sequence events that are far apart.

A meaningful relationship between sequence events separated by many tokens.

A meaningful relationship between musical events separated by many tokens or seconds.

A distribution with a few very common categories and many categories containing relatively few examples.

Musical relationships that extend across relatively long spans, such as phrases, motifs, and sections.

A failure mode where music becomes stuck repeating a short pattern with little development.

A parameter-efficient method that learns low-rank updates to selected frozen weight matrices.

The inner dimension controlling the size and capacity of a LoRA update.

A numerical measure of prediction error used to guide model updates.

A plot or sequence showing loss values across training progress.

A function converting predictions and targets into a numerical optimisation objective.

Changing how strongly errors from selected examples or groups affect the training objective.

Compression that removes some information to produce a smaller representation.

Adjustment of playback or file gain so stimuli have similar measured programme loudness.

Adjusting gain so measured loudness moves toward a chosen target.

A matrix update represented as the product of two smaller matrices.

A unit commonly used to express loudness measured with an algorithm based on ITU-R BS.1770.

The words intended to be sung in a song.

Guiding vocal music generation with words intended to be sung.

A tendency for an output requested in a low-resource or niche style to move toward patterns from more common styles.

A probabilistic sequence model whose next prediction depends on a limited recent state history.

Generation in which selected positions or regions are hidden and reconstructed from visible context and conditions.

A selected part of a representation hidden from the model and generated or replaced.

Rights and permissions connected with a particular sound recording.

The practical ability to understand, direct, reject, inspect, reverse, or escalate important system actions.

A skew caused by using a metric or proxy that does not measure the intended concept equally well across groups.

Resemblance between pitch and timing patterns under a defined representation and permitted transformations.

Guiding generation with a representation of melodic pitch movement and timing.

An attempt to determine whether a candidate example was included in model training.

An attempt to infer whether an item was part of a model's training data from observable behaviour.

Retention of information specific to one or more training examples.

Retention of example-specific training information that can influence or be reproduced by a model.

Information that describes, organises, or helps analyse a record.

Information that describes, identifies, organises, or gives context to a music record.

Improving a measured score without equivalent improvement in the intended real-world objective.

The consistency of a metric under equivalent defined conditions.

The degree to which a metric provides evidence about its intended property and decision.

The examples processed in one forward and backward pass on one worker.

A technical standard for sending and storing musical performance instructions as timed events.

A technical standard for representing musical performance instructions as timed events.

A protocol and file format representing musical performance messages and timing.

A timed instruction such as note on, note off, program change, pitch bend, or controller movement.

A subset of the training dataset used to estimate an update.

The smallest program or pipeline that still demonstrates a failure.

A relevant audible property that a caption or annotation leaves out.

An expected trainable parameter has no calculated gradient after backpropagation.

Training that uses more than one numerical precision across operations.

Training that uses more than one numerical precision across operations and stored values.

Compatibility between generated stems and the combined full mix.

A failure where a generator produces a narrow range of outputs relative to the intended distribution.

The range and complexity of functions or patterns a model can represent.

Documentation describing a model's purpose, intended use, evaluation, limitations, risks, and context.

A persistent saved record containing model and optional training state.

One forward computation of the denoising network during sampling.

Distributing different model components or tensor operations across devices.

An immutable identifier for a specific deployable model artifact and associated configuration.

Audio represented with one channel.

A short recognisable musical idea that can return or develop across a piece.

The features that allow a motif to remain recognisable across repetitions or variations.

The return of a recognisable motif later in a piece.

Similarity between music-audio and natural-language representations produced by MuLan.

An annotation setup where one example can receive several labels at the same time.

Several token sequences representing related information, such as multiple neural codec codebooks.

Conditioning music on several input types such as text, audio, image, video, or symbolic events.

A recording project in which sources are preserved on separate synchronized tracks.

A controlled subjective method for assessing intermediate levels of audio quality using multiple stimuli and anchors.

A multi-stimulus audio-quality method using a reference, hidden reference, and anchors.

Generating a selected region while conditioning on surrounding or preserved material.

The underlying music and any lyrics, considered separately from a particular recorded performance.

A sequence model estimating the next event from the previous fixed number of events.

A floating-point value representing an undefined or invalid numerical result.

Two items that differ in presentation while preserving much of the same underlying content.

A different file that contains the same or nearly the same source audio after changes such as re-encoding, trimming, or gain adjustment.

An item whose embedding is among the closest to a query embedding according to a chosen similarity measure.

Retrieval of the most similar training items for each generated sample under selected metrics.

A loss related to the negative logarithm of the probability assigned to the observed target.

Two training examples that should not be treated as a match.

An audio codec that uses neural networks to learn how to compress and reconstruct sound.

A learned encoder, quantiser, and decoder that compress audio into tokens and reconstruct it.

Estimating which event is likely to occur after the available context.

A licence condition restricting the sharing of adapted material.

A diffusion objective where the model estimates the noise added to a clean example.

A rule determining the amount of noise associated with each diffusion timestep.

A licence condition limiting permission to uses that meet the licence's noncommercial definition.

Difference between an output and a selected comparison set under a defined representation or metric.

Half the sample rate and the highest frequency that can be represented without aliasing under ideal sampling conditions.

Storage that manages files as objects with data, identifiers, and metadata.

The ability to understand system behavior through emitted telemetry such as traces, metrics, and logs.

A structured set of concepts and relationships used to describe a domain.

A rule translating an abstract concept into an observable or calculable measure.

Training behaviour characterised by erratic loss, large updates, non-finite values, or sensitivity that prevents reliable progress.

The chain connecting model inputs and targets to loss, gradients, optimiser state, and parameter updates.

An algorithm that converts gradients into parameter updates.

Persistent values maintained by an optimiser across parameter updates.

A change in judgement caused by stimulus presentation position or sequence.

A cryptographic digest used to identify the exact bytes of a source file.

The degree to which output provides a new result rather than unacceptable reproduction of existing material.

A failure caused by an allocation request exceeding the available usable memory.

Extending a known recording beyond one of its existing boundaries.

Improving on training examples without equivalent improvement on unseen examples.

A generalisation failure where training performance improves more than relevant held-out performance.

Audio content shared by two neighbouring segments.

Showing selected examples or groups more frequently during training.

A record showing which positions contain real audio and which positions were added as padding.

An indicator showing which sequence positions contain valid data.

A measure comparing one evaluated item with one corresponding reference or condition.

A test where two items are compared directly on a defined criterion.

A known target corresponding directly to an evaluated output.

Variation estimated from distances among pairs of generated outputs.

Predicting or refining several token positions during the same decoding iteration.

A trainable numerical value inside a model.

Memory used to store model weights and other model-state tensors.

A change to the model's learned numerical parameters during training.

A numerical change applied to trainable model parameters.

Adapting a pretrained model by training only a small subset or small set of added parameters.

Updating selected existing layers while freezing the rest of a pretrained model.

A generated sequence matching a defined proportion or region of a reference.

A condition where earlier generated choices change the later direction and available options.

The greatest amount of memory used or reserved during a measured period.

An agreement documenting permission from a performer for specified uses.

Generation adapted to a user's or group's preferences, history, profile, or immediate context.

A grid used to display and edit notes across time and pitch.

A grid representation showing active pitches across discrete time positions.

A previously working behaviour that becomes incorrect after a system change.

The perceptual quality that makes a sound seem higher or lower.

A documented result showing whether and how an asset may be used under a named policy version.

Two training examples that should be treated as related, such as a music clip and its matching description.

Transformations applied after model generation to create validated deliverable assets.

A difference large enough to matter for the intended decision.

A model input calculated before training and stored for repeated use.

The scores or probabilities produced by a model for possible outputs.

A listener's relative or absolute liking for an item.

A model estimating which outputs a person or population is likely to prefer.

The sequence of tokens appearing before the position currently being predicted.

A time-limited URL granting a specific object-storage operation.

A model whose parameters have already been trained on an earlier dataset or objective.

An initial training stage intended to learn broad representations or generative behaviour.

A set of probabilities assigned across possible outcomes, adding up to one.

A record of changes applied to data, such as resampling, segmentation, cleaning, or annotation.

A record of every change applied to a source file when creating a training derivative.

Measuring execution time, memory allocation, communication, and input-pipeline activity.

An objective estimate of perceived loudness calculated across an audio programme.

User-supplied information requesting or describing a generated result.

The degree to which generated music follows the requested description or condition.

The extent to which generated output matches a text description.

A gradual movement away from the style, content, or attributes requested by the original condition.

An immutable identifier for the rules and transformations converting user input into model conditions.

Information about where data came from, how it was processed, and what permissions or restrictions apply.

Information about where a record or annotation came from and how it was created or changed.

Recorded information about the origin, history, ingredients, and transformations of an asset.

A measurable quantity used as an indirect indicator of another property.

A status where copyright restrictions do not apply to a work in the relevant jurisdiction, while separate rights in related material may still need review.

A mandatory condition that must be met regardless of the overall score.

The process of replacing a value with a choice from a limited set of available values.

A separate review area for records that failed validation or need a human decision.

Kept out of training while missing, conflicting, disputed, or uncertain evidence is reviewed.

A segment whose start position is selected randomly from a longer recording.

Assigning initial parameter values using a randomised numerical scheme before training.

A value used to initialise a repeatable sequence of random numbers.

A value used to initialise a pseudo-random process for more consistent repetition.

Assignment of presentation order or condition position using a random procedure.

A written or audible reference defining what a score represents.

A written or audible reference explaining what a score represents.

A defined set or continuum of response values used to express a judgement.

Audio represented directly as a sequence of sampled signal values rather than as notes or another higher-level description.

Processing time divided by generated audio duration.

Continuous or rapidly updated generation responding within a latency suitable for live interaction.

A tendency for recent context to influence a prediction more strongly than distant information.

The audio produced by a decoder from a compressed representation.

The degree to which an output preserves a known reference signal or its perceptually important properties.

The proportion and types of relevant material represented in a comparison index.

A collection of real or target examples used as a comparison distribution or benchmark.

A collection representing the target distribution used for comparison.

Distortion of an evaluation caused by the composition or limitations of the comparison collection.

A test designed to detect whether a previously corrected or working behaviour becomes incorrect again.

A strategy intended to improve behaviour on unseen data rather than only minimise training error.

Earlier or representative examples mixed into later training to help preserve previous capabilities.

An attention method that represents relationships between positions using their relative distance.

The consistency with which a system meets defined requirements.

Repetition of a trial to estimate consistency in one participant's responses.

The return of recognisable musical material with selected changes.

A comparison of generated material with training or reference material to identify possible reproduction.

The presence and variety of a group, category, or musical characteristic within a dataset.

A compressed representation that limits how much information can pass into later stages.

The degree to which an experiment can be repeated using documented data, code, configuration, and randomness controls.

The degree to which an experiment and its results can be reconstructed or repeated.

A failure that can be triggered consistently under recorded conditions.

A versioned schema defining accepted inputs, defaults, constraints, and expected outputs.

Converting a digital audio signal from one sample rate to another.

An implemented method evaluated under reported experimental conditions.

GPU memory retained by the framework allocator for current or later tensor allocations.

A quantisation method where several codebooks progressively encode the remaining representation error.

A method that uses several codebooks in sequence so later codebooks can encode information missed by earlier ones.

A checkpoint containing enough state to continue an interrupted training run.

The learned process that transforms a noisy representation toward generated data.

A status used when evidence exists but a policy or legal interpretation remains uncertain.

Withdrawal or termination of authorization for future access or use under defined conditions.

Guiding generation with beat, onset, groove, or drum information.

A person or organisation that controls one or more legal rights in an asset.

Restoring a previously approved model or pipeline version after a failed change.

The underlying mechanism that produces the observed failure.

A check that data can be encoded and decoded without unexpected changes.

A test that encodes data into an intermediate representation and decodes it for comparison with the source.

Differences between experimental runs caused by random initialisation, batch order, sampling, or numerical behaviour.

One numerical measurement of an audio signal at a particular moment.

The number of audio samples taken per second, measured in hertz.

The number of audio samples represented each second.

A measure calculated for one generated item.

The algorithm that converts model predictions into updates during diffusion generation.

Selecting an output according to a probability distribution.

One model evaluation and representation update during generation.

A value that changes how likely a record is to be selected during training.

A small test verifying that part of the training pipeline behaves as expected.

A training approach that gradually mixes model-generated prefixes into a previously guided sequence-training process.

The assets, activities, products, period, and conditions covered by a person's agreement.

The defined purposes, users, projects, territories, media, duration, and transformations permitted under an authorization.

A region of music with a recognisable structural role or identity.

The musical properties that make a section recognisable when it returns.

The return of a recognisable section later in a piece.

The musical passage or change connecting one structural section with another.

Consistency of required behaviour across random seeds for the same condition.

Dividing a longer recording into shorter time ranges used as separate examples.

A skew created by the rules or practical limits used to choose which records enter a dataset.

A comparison showing how similar different time regions of one piece are to one another.

A matrix comparing representations of different time regions within one piece.

Compatibility between the high-level meaning of generated audio and its condition.

A model that predicts or processes high-level learned content tokens.

A learned discrete token intended to preserve relatively high-level content or long-term structure.

A model designed to process or generate ordered data.

A fixed or bounded section of a longer token sequence used as one model input or target.

Generation where later outputs must wait for earlier outputs to be selected.

A target for a measurable aspect of service reliability or performance over a defined period.

A measure calculated over a collection of generated or reference items.

A test where a new version processes copied requests without controlling the user-visible response.

Dividing model, optimiser, data, or checkpoint state across workers or files.

A licence condition requiring certain adaptations to use the same or a compatible licence.

A method that divides a signal into short sections and measures the frequency content of each section.

A measurement comparing the strength of the structured signal with the strength of noise.

A level below which a frame is treated as silent by a chosen detection rule.

A measured or perceived relationship under a specified representation and comparison rule.

A function measuring resemblance between two musical or audio representations.

A metric value used to flag or classify a pair under an evaluation protocol.

Perceived or measured resemblance between vocal performances with respect to performer identity cues.

A test where one item is presented and rated without an immediate alternative.

An update not applied because the system detected invalid gradients or another update condition.

A strategy that moves a fixed-size context across a longer sequence.

A limited study designed to answer one important question with modest data, compute, and implementation complexity.

A function that converts logits into a probability distribution.

The large-scale arrangement of musical sections across a piece.

A fixation of a series of sounds, such as a particular recorded performance.

A travelling pattern of pressure changes moving through a medium such as air or water.

The original or upstream file from which a dataset record or derivative was created.

A skew caused by relying heavily on particular platforms, archives, labels, or contributors.

The degree to which a dataset depends on a small number of platforms, archives, or catalogues.

The original label exactly as it appeared in the source data.

Unwanted presence of one source within another separated or generated stem.

Resemblance involving features sufficiently specific to connect an output with an identifiable reference.

A vector representation trained to capture voice characteristics useful for distinguishing speakers.

Testing whether a recording is consistent with a claimed speaker identity.

A difference calculated between time-frequency representations of two audio signals.

A representation that shows how a signal's frequency content changes over time.

Applying diffusion to a time-frequency representation of audio.

A process where a problem in an earlier stage affects the inputs and outputs of later stages.

An isolated or grouped component of a multitrack mix.

Generation that represents and controls multiple musical sources separately.

Audio represented with two channels, conventionally left and right.

An optimiser family using gradients estimated from individual examples or mini-batches.

A predefined condition that ends or pauses an experiment.

A condition that ends generation, such as a stop token, duration, or maximum sequence length.

Rules governing how stored objects transition, expire, or remain under retention.

A storage category selected according to speed, durability, and access requirements.

Control information specifying musical form, sections, lyrics, chords, or other large-scale organisation.

A representation of form, phrases, sections, or relationships used to guide lower-level generation.

A graph connecting musical regions through relationships such as repetition, variation, or development.

A mechanism preserving or retrieving earlier information according to musical roles.

Using a reference signal to guide broad musical or acoustic characteristics while generating new content.

Resemblance in broad musical or production characteristics shared across a category or body of work.

Measurement based on reported human perception, opinion, or preference.

A predefined condition used to determine whether an experiment achieved its objective.

Increasing the detail or resolution of an existing representation.

A representation of musical events and instructions rather than recorded audio samples.

A hardware or software instrument that creates sound and can respond to MIDI instructions.

A caption generated automatically by a model rather than written directly by a human annotator.

A measure of behaviour across repeated requests, prompts, failures, or operating conditions.

A short descriptive label such as piano, jazz, speech, or percussion.

A short label used to describe or categorise one property of a recording.

A category with relatively few training examples.

Latency experienced by the slowest portion of requests, commonly represented by a high percentile.

The expected result that a model is trained to predict.

The correct output from the training data that the model is asked to predict.

The desired mix of examples or training exposure chosen for a model's intended use.

An organised system of categories and relationships used to classify items.

Training a sequence model with the correct earlier tokens from the dataset.

A generation setting that makes a sampling distribution more concentrated or flatter.

Guiding generation with a target pace, commonly represented in beats per minute.

Consistency of estimated rhythmic pace across a piece.

A mapping between a condition and the times or sequence positions where it applies.

Mapping event times onto a fixed rhythmic grid or finite timing vocabulary.

Held-out data reserved for final evaluation after model choices are complete.

Guiding generation with a numerical representation of written language.

An encoder that turns text into a learned representation such as an embedding.

The task of generating music that follows a written description.

The task of generating music that follows a written description.

The sound quality that helps distinguish two instruments or voices even when they play the same pitch.

A value marking when an event begins, ends, or occurs within a recording.

A numerical representation that tells the model which diffusion noise level it is processing.

A numerical representation of timing information supplied as a model condition.

The smallest timing interval represented by a symbolic encoding.

Repeated training on a small fixed dataset to verify that the pipeline can reduce its training error.

One discrete unit in the sequence processed or generated by a model.

A rule for arranging tokens from several streams into an autoregressive prediction order.

The number of tokens represented or generated per second of audio or sequence time.

The number of model tokens associated with one unit of audio time.

A sequence of token IDs ordered over time.

Consistency or intentional evolution of pitch-class and tonal information over time.

The fraction of examples or hours contributed by the largest artist, source, or category.

Sampling only from a fixed number of the highest-scoring token candidates.

Sampling from the smallest set of tokens whose combined probability reaches a chosen threshold.

A difference between the sequence histories or conditions used during training and free generation.

The number of parameters that receive gradients and optimiser updates.

The process of adjusting model parameters to reduce a chosen loss across training examples.

A predefined limit on updates, compute, time, data exposure, or resources.

One structured record presented to a model during training.

Training relevant model parameters from random or newly constructed initial values.

The hardware, storage, network, software, scheduling, and tracking systems used to run model experiments.

Loss calculated on examples used in the optimisation process.

A versioned record identifying the examples or source groups used in a model-training run.

Data used to calculate losses and update model parameters.

One batch iteration that usually includes one optimiser update.

The amount of training work completed per unit of time.

Generation or recovery of recognisable training examples from a trained model.

Elicitation of model output that reproduces identifiable training information.

Whether a particular item was included in data used to update model parameters.

Reusing features learned by a pretrained model for a new task.

Reusing knowledge from a pretrained model for a new task, dataset, or domain.

A record of processing steps such as decoding, resampling, segmentation, annotation, and filtering.

A comparison property that treats a melody shifted uniformly in pitch as equivalent or similar.

An estimate of the highest signal peak, including peaks that may occur between stored samples.

An estimate of the maximum level of the reconstructed continuous audio signal.

A dimensionality reduction method often used to display high-dimensional data in two or three dimensions.

Lyrics supplied without exact timing for each word, syllable, or phoneme.

Failure to model important patterns even in the training data or target domain.

A condition where a relevant group or combination appears too rarely or narrowly for the intended use.

Reducing how often common examples or groups are used during training.

Measurement of cost and value per defined product unit.

A label or claim that cannot be justified by the audio, trusted metadata, or another documented source.

A model that predicts a denser or more detailed representation from a coarser one.

The degree to which a result helps a defined user complete an intended task.

The proportion of generation attempts satisfying predefined technical output requirements.

The proportion of generation attempts producing outputs that satisfy defined technical requirements.

A validation design or process that does not reliably estimate intended generalisation performance.

Loss calculated on held-out validation data without parameter updates.

Indirect adaptation to validation examples through repeated experiment decisions.

Evaluation on held-out data without parameter updates.

Held-out data used for model selection and experiment decisions without parameter updates.

A new version that preserves selected properties of existing music while changing others.

An ordered list of numbers. In an embedding, the numbers act as coordinates in a learned space.

A method that replaces a continuous vector with a selected vector from a codebook.

A MIDI value commonly used to represent how strongly a note was played.

The period during which a queue message is hidden after one worker receives it.

The complete set of token types the model can predict.

A model that converts an acoustic representation such as a mel-spectrogram into a waveform.

A vector-quantised variational autoencoder that learns discrete latent codes and a decoder.

A serving process with required model artifacts already loaded and ready for inference.

The degree to which a watermark remains detectable after defined transformations or attacks.

A representation of how an audio signal changes over time.

Applying diffusion directly to time-domain audio sample values.

A label that says an event appears somewhere in a clip without giving its exact timing.

A label stating that a property occurs somewhere in a clip without giving its exact time.

An optimisation regulariser that discourages large parameter values through a decay term or decoupled update.

A composite score combining dimensions according to predefined importance weights.

A checkpoint containing model parameters without the complete optimisation state.

Assessment that examines the complete generated timeline rather than isolated excerpts.

Condition alignment calculated separately over several time regions.

An action ending or changing future reliance on a consent or permission under the applicable agreement or legal framework.

Evaluation of how a person and system complete a task rather than evaluation of one isolated output.

A calculated gradient whose values are zero or effectively negligible.

Classifying an item using new text labels or descriptions without training a separate classifier on labelled examples for every class.

Applying a pretrained model to a task or input without task-specific parameter training.