Localizing Audio Definition Model metadata is a matter of preserving a playable audio scene while making its language versions and descriptive labels intelligible to different audiences. ADM, specified in ITU-R BS.2076, expresses how an audio programme relates to content components, objects, channel layouts and audio tracks. BW64 provides an important way to carry ADM metadata with actual audio material. Translating an object name is one small part of the task; preserving the complete relationships is the harder and more important part.
If you need to translate ADM XML, localize BW64 audio programmes, distinguish audioProgramme and audioContent language tags, manage complementary dialogue objects, or validate axml and chna track mapping for immersive broadcast delivery, this guide explains a safe approach. It discusses programme selection, object identity, language-dependent dialogue, shared ambience, loudness, spatial rendering, playback configurations and validation under the applicable standards and production profiles.
This guide serves audio engineers, broadcasters, post-production teams, accessibility specialists, localization vendors, streaming delivery teams and learners worldwide. It does not prescribe one codec, one brand of renderer or one target-market naming convention. The current published framework includes ITU-R BS.2076-3, approved in February 2025, and ITU-R BS.2088-2 for BW64, approved in November 2025; implementation profiles may constrain the broader model. Any real delivery must follow the recipient’s agreed versions and profile rather than an assumed universal workflow.
The first principle: a language option must be an audible option
Do not accept a metadata file merely because it announces English, German, Arabic or another language in a descriptive field. Follow every advertised presentation through its programme, content and object references to the actual audio tracks. Then listen. This is the defining distinction between localization of an ADM graph and cosmetic translation of its labels.
Map the ADM graph before translating names
The difference between an audio file and ADM metadata
ADM describes audio content, relationships and playback-relevant properties; the audio samples themselves are carried by an associated media representation. Think in paths through the audio graph. The listener does not hear a text field; the listener hears the samples selected and rendered through a chain of references. Names help operators understand those paths, but they cannot substitute for the technical relationships. This distinction governs which properties may be translated.
A programme delivers English and French dialogue possibilities while sharing the same music and effects. The description must identify each meaningful mix. For the acceptance record, identify which audioProgramme and component IDs are involved and what the receiver should hear. Record the relevant language choices and graph links so that a second engineer can reproduce the test rather than trusting an assertion that the XML was translated.
Treating ADM as nothing more than a multilingual list of track names misses the graph that tells a renderer which tracks make each programme. This defect can remain undetected in a text-only inspection because the metadata may still parse and the media may still open. The real test must check both the intended model and its audible consequences.
Identify both the carrier format and the ADM graph, then verify that each intended presentation resolves to available audio essence. Record source and target values, the applicable standard or production profile, the rendering tool and any exception. Use the same acceptance conditions for the original and localized versions so differences can be explained precisely.
audioProgramme as a listener-facing presentation
An audioProgramme represents a whole programme or a version of one, and can reference multiple audioContent components. A localized audio programme can be grammatically labeled correctly and still be the wrong programme. Its language metadata, linked content, object relationships and rendering behavior must agree. The appropriate review therefore follows the user’s choice all the way to audible output.
An international sports broadcast has a programme for one language and another programme for a second language, with common stadium ambience. A side-by-side comparison should distinguish allowed display-text changes from invariant links, timing and sample mappings. This prevents the team from confusing a metadata correction with a new audio mix. When a different mix is needed, explicitly open a mastering task with its own evidence.
Changing only an audioProgrammeName can leave both programme versions pointing to the same dialogue. The listener then hears the wrong language despite the displayed label. This defect can remain undetected in a text-only inspection because the metadata may still parse and the media may still open. The real test must check both the intended model and its audible consequences.
For each programme, enumerate referenced audioContent IDs and render the resulting selection while checking the declared language. Record source and target values, the applicable standard or production profile, the rendering tool and any exception. Use the same acceptance conditions for the original and localized versions so differences can be explained precisely.
audioContent as a meaningful component
audioContent describes a component such as dialogue, commentary, music or ambience and can carry language and loudness-related information. The broad ADM model is useful because it separates what a programme means from how its audio is structured. That flexibility demands care: a translated value may be descriptive, a code may be structural and a numeric field may affect the listener’s spatial or temporal experience.
A documentary programme contains one narration audioContent and separate language-independent content for score and sound effects. The receiving system’s behavior is authoritative for interoperability. A generic XML reader can confirm syntax but cannot demonstrate that a playout pipeline understands complementary objects, gain limits or BW64 channel allocation. Test with the intended renderer and document limitations.
Translating the name Music into another language does not transform music into dialogue; incorrectly tagging it as language-specific can break selection. This defect can remain undetected in a text-only inspection because the metadata may still parse and the media may still open. The real test must check both the intended model and its audible consequences.
Verify the semantic role, language metadata where applicable and every referenced audioObject in the recipient’s interpretation. Record source and target values, the applicable standard or production profile, the rendering tool and any exception. Use the same acceptance conditions for the original and localized versions so differences can be explained precisely.
audioObject as the link to actual audio
audioObject connects content intent to format and track references. Its relationships can include nesting and complementary alternatives. Multilingual delivery is a design question as well as a language question. Decide whether the recipient wants independent fixed presentations, interactive alternatives or separate exported masters. Each path has different metadata and testing needs. A translator should not silently choose an interaction model on behalf of a broadcaster.
A French narrator object points to the correct track UID while a background object points to a shared effects bed. Where there is disagreement, investigate it at the correct layer: language label, language code, graph reference, channel mapping, sample content or renderer configuration. Altering a translated title cannot repair a missing track. A structured defect report shortens the route to a genuine fix.
An object name alone cannot guarantee the right voice is played. References to a different track or pack can silently substitute material. This defect can remain undetected in a text-only inspection because the metadata may still parse and the media may still open. The real test must check both the intended model and its audible consequences.
Resolve object IDs to their track UIDs and pack formats and listen to the selected mix, especially at language changes. Record source and target values, the applicable standard or production profile, the rendering tool and any exception. Use the same acceptance conditions for the original and localized versions so differences can be explained precisely.
ID attributes and ID references
IDs such as APR, ACO, AO and ATU identify distinct ADM elements. IDRef relationships form the graph’s internal connections. The most reliable proof is reproducible playback. Editors can compare titles, identifiers and language tags, but only an actual render can show the correct narrator, shared ambience, spatial motion and accessible selection. Treat both the data graph and the rendered experience as deliverables.
A graph contains audioProgrammeID APR_1001 and links to audioContentID ACO_1001 and audioObjectID AO_1001. Maintain a small representative test suite with at least two languages, shared audio, a spatial event and a known default selection. Run it before and after localization. It is easier to catch a graph regression with a focused example than at the end of a long programme.
Translating or renumbering an ID without updating every reference can orphan part of the graph or point to another object. This defect can remain undetected in a text-only inspection because the metadata may still parse and the media may still open. The real test must check both the intended model and its audible consequences.
Compare ID sets and directed reference edges; reject unresolved links, unintended duplicates and unexpected changes to element identity. Record source and target values, the applicable standard or production profile, the rendering tool and any exception. Use the same acceptance conditions for the original and localized versions so differences can be explained precisely.
AudioFormatExtended and version context
audioFormatExtended acts as a parent for ADM elements and may carry version context. The receiving renderer may support a specified subset. Think in paths through the audio graph. The listener does not hear a text field; the listener hears the samples selected and rendered through a chain of references. Names help operators understand those paths, but they cannot substitute for the technical relationships. This distinction governs which properties may be translated.
A studio has an ADM file authored to a newer recommendation than an older broadcast playout chain can accept. For the acceptance record, identify which audioProgramme and component IDs are involved and what the receiver should hear. Record the relevant language choices and graph links so that a second engineer can reproduce the test rather than trusting an assertion that the XML was translated.
Editing the declared version string is not a structural conversion. Unsupported features will not become supported through a cosmetic label change. This defect can remain undetected in a text-only inspection because the metadata may still parse and the media may still open. The real test must check both the intended model and its audible consequences.
Validate the actual ADM version and delivery profile with the recipient before translation, and test unsupported features explicitly. Record source and target values, the applicable standard or production profile, the rendering tool and any exception. Use the same acceptance conditions for the original and localized versions so differences can be explained precisely.
Connect the graph to BW64 media correctly
The BW64 container and long-form exchange
BW64 builds on the WAVE/RF64 family to support larger audio programmes and metadata-bearing multichannel exchange. A localized audio programme can be grammatically labeled correctly and still be the wrong programme. Its language metadata, linked content, object relationships and rendering behavior must agree. The appropriate review therefore follows the user’s choice all the way to audible output.
A production exchanges a long immersive programme with more channels and metadata than a simple stereo WAV handoff. A side-by-side comparison should distinguish allowed display-text changes from invariant links, timing and sample mappings. This prevents the team from confusing a metadata correction with a new audio mix. When a different mix is needed, explicitly open a mastering task with its own evidence.
A generic WAV editor may discard metadata chunks or fail on larger-file structures. Playback success does not imply the ADM scene survived. This defect can remain undetected in a text-only inspection because the metadata may still parse and the media may still open. The real test must check both the intended model and its audible consequences.
Use BW64-aware tools and inspect the actual container, size records, audio channels and metadata chunks before and after the edit. Record source and target values, the applicable standard or production profile, the rendering tool and any exception. Use the same acceptance conditions for the original and localized versions so differences can be explained precisely.
The axml metadata chunk
The axml chunk carries ADM information in an XML representation when embedded in the relevant BW64 workflow. The broad ADM model is useful because it separates what a programme means from how its audio is structured. That flexibility demands care: a translated value may be descriptive, a code may be structural and a numeric field may affect the listener’s spatial or temporal experience.
A post-production system reads the embedded axml to determine the language choices and spatial rendering instructions for a programme. The receiving system’s behavior is authoritative for interoperability. A generic XML reader can confirm syntax but cannot demonstrate that a playout pipeline understands complementary objects, gain limits or BW64 channel allocation. Test with the intended renderer and document limitations.
Replacing the axml block with a shortened translated summary loses the complete model and can make a valid audio file semantically incomplete. This defect can remain undetected in a text-only inspection because the metadata may still parse and the media may still open. The real test must check both the intended model and its audible consequences.
Parse the XML and retain the intended ADM element hierarchy, IDs, timing and relationships while changing only approved human-language fields. Record source and target values, the applicable standard or production profile, the rendering tool and any exception. Use the same acceptance conditions for the original and localized versions so differences can be explained precisely.
The chna channel-allocation chunk
The chna chunk ties file tracks to ADM identifiers including track UIDs and format references. It is a crucial bridge between graph and samples. Multilingual delivery is a design question as well as a language question. Decide whether the recipient wants independent fixed presentations, interactive alternatives or separate exported masters. Each path has different metadata and testing needs. A translator should not silently choose an interaction model on behalf of a broadcaster.
A multichannel file has physical track positions that a renderer must associate with the correct audio objects. Where there is disagreement, investigate it at the correct layer: language label, language code, graph reference, channel mapping, sample content or renderer configuration. Altering a translated title cannot repair a missing track. A structured defect report shortens the route to a genuine fix.
If a translator or converter changes the mapping, a dialogue object might be rendered from an effects track. An attractive label cannot fix that. This defect can remain undetected in a text-only inspection because the metadata may still parse and the media may still open. The real test must check both the intended model and its audible consequences.
Check chna track indexes and identifiers against axml and the actual media channel layout with a specialist parser. Record source and target values, the applicable standard or production profile, the rendering tool and any exception. Use the same acceptance conditions for the original and localized versions so differences can be explained precisely.
audioTrackUID relationships
An audioTrackUID identifies a track representation used by objects and in the allocation mapping. It is not a phrase for localization. The most reliable proof is reproducible playback. Editors can compare titles, identifiers and language tags, but only an actual render can show the correct narrator, shared ambience, spatial motion and accessible selection. Treat both the data graph and the rendered experience as deliverables.
Two narrations can use distinct track UIDs, allowing a programme to choose the proper dialogue without duplicating all common audio. Maintain a small representative test suite with at least two languages, shared audio, a spatial event and a known default selection. Run it before and after localization. It is easier to catch a graph regression with a focused example than at the end of a long programme.
Replacing an ATU reference with a translated track label severs the object from the sample-bearing route. This defect can remain undetected in a text-only inspection because the metadata may still parse and the media may still open. The real test must check both the intended model and its audible consequences.
Preserve UID values and validate that each referenced UID resolves through the carrier’s allocation information to the intended samples. Record source and target values, the applicable standard or production profile, the rendering tool and any exception. Use the same acceptance conditions for the original and localized versions so differences can be explained precisely.
Common definitions and custom definitions
ADM implementations can reference common format definitions as well as locally defined formats. Their identities have structured meaning. Think in paths through the audio graph. The listener does not hear a text field; the listener hears the samples selected and rendered through a chain of references. Names help operators understand those paths, but they cannot substitute for the technical relationships. This distinction governs which properties may be translated.
A common front-left speaker definition may be reused while a custom object trajectory is defined within the delivered metadata. For the acceptance record, identify which audioProgramme and component IDs are involved and what the receiver should hear. Record the relevant language choices and graph links so that a second engineer can reproduce the test rather than trusting an assertion that the XML was translated.
Changing standard format identifiers or confusing common references with custom ones can make a valid-looking track impossible to render correctly. This defect can remain undetected in a text-only inspection because the metadata may still parse and the media may still open. The real test must check both the intended model and its audible consequences.
Resolve common-definition references against the applicable standard and ensure custom definitions exist where required. Record source and target values, the applicable standard or production profile, the rendering tool and any exception. Use the same acceptance conditions for the original and localized versions so differences can be explained precisely.
The audio samples and track order
Physical sample order is not the same as human naming order. Channel indexing matters even when different audio objects have descriptive labels. A localized audio programme can be grammatically labeled correctly and still be the wrong programme. Its language metadata, linked content, object relationships and rendering behavior must agree. The appropriate review therefore follows the user’s choice all the way to audible output.
A file contains ambience before narration in its interleaved channel order but a localized UI lists narration first. A side-by-side comparison should distinguish allowed display-text changes from invariant links, timing and sample mappings. This prevents the team from confusing a metadata correction with a new audio mix. When a different mix is needed, explicitly open a mastering task with its own evidence.
Sorting tracks alphabetically by translated name can destroy the mapping between metadata and actual samples. This defect can remain undetected in a text-only inspection because the metadata may still parse and the media may still open. The real test must check both the intended model and its audible consequences.
Check channel index, waveform identity and ADM graph correspondence; test audible content from the rendered selection. Record source and target values, the applicable standard or production profile, the rendering tool and any exception. Use the same acceptance conditions for the original and localized versions so differences can be explained precisely.
Build safe multilingual programme choices
Programme language and content language
Language attributes can describe a whole programme presentation and individual components. Their meaning depends on how alternatives are modeled. The broad ADM model is useful because it separates what a programme means from how its audio is structured. That flexibility demands care: a translated value may be descriptive, a code may be structural and a numeric field may affect the listener’s spatial or temporal experience.
An English and a German presentation share effects but select different dialogue content. Programme and content languages help a player identify the choice. The receiving system’s behavior is authoritative for interoperability. A generic XML reader can confirm syntax but cannot demonstrate that a playout pipeline understands complementary objects, gain limits or BW64 channel allocation. Test with the intended renderer and document limitations.
Tagging the common score with a dialogue language or leaving two different narration versions ambiguously identified may mislead automated selection. This defect can remain undetected in a text-only inspection because the metadata may still parse and the media may still open. The real test must check both the intended model and its audible consequences.
For each target presentation, document which language-bearing components it includes and check consistency of the language declarations. Record source and target values, the applicable standard or production profile, the rendering tool and any exception. Use the same acceptance conditions for the original and localized versions so differences can be explained precisely.
Noninteractive multilingual programmes
One supported approach to language versions is defining separate audioProgrammes whose referenced components produce the desired language presentation. Multilingual delivery is a design question as well as a language question. Decide whether the recipient wants independent fixed presentations, interactive alternatives or separate exported masters. Each path has different metadata and testing needs. A translator should not silently choose an interaction model on behalf of a broadcaster.
A broadcaster prepares an Italian programme and an English programme that reuse the same music and ambience. Where there is disagreement, investigate it at the correct layer: language label, language code, graph reference, channel mapping, sample content or renderer configuration. Altering a translated title cannot repair a missing track. A structured defect report shortens the route to a genuine fix.
If both programme entries accidentally reference the Italian narrator, the labels can look correct while the English delivery is wrong. This defect can remain undetected in a text-only inspection because the metadata may still parse and the media may still open. The real test must check both the intended model and its audible consequences.
Render each programme independently and verify narration, shared beds, start and end boundaries and final output metadata. Record source and target values, the applicable standard or production profile, the rendering tool and any exception. Use the same acceptance conditions for the original and localized versions so differences can be explained precisely.
Interactive complementary audio objects
Another ADM approach represents mutually exclusive alternatives with complementary audioObjects, allowing an appropriate presentation selection. The most reliable proof is reproducible playback. Editors can compare titles, identifiers and language tags, but only an actual render can show the correct narrator, shared ambience, spatial motion and accessible selection. Treat both the data graph and the rendered experience as deliverables.
A viewer switches between languages during playback, while the same programme retains its common effects and music. Maintain a small representative test suite with at least two languages, shared audio, a spatial event and a known default selection. Run it before and after localization. It is easier to catch a graph regression with a focused example than at the end of a long programme.
Copying a language object without correct complementary relationships may cause both narrators to play at once or hide a valid option. This defect can remain undetected in a text-only inspection because the metadata may still parse and the media may still open. The real test must check both the intended model and its audible consequences.
Check the complementary-object structure against the intended interaction model and audition every supported language choice. Record source and target values, the applicable standard or production profile, the rendering tool and any exception. Use the same acceptance conditions for the original and localized versions so differences can be explained precisely.
Shared ambience and music beds
Language-independent content is often reused across versions. Shared components reduce unnecessary duplication and help keep the experience consistent. Think in paths through the audio graph. The listener does not hear a text field; the listener hears the samples selected and rendered through a chain of references. Names help operators understand those paths, but they cannot substitute for the technical relationships. This distinction governs which properties may be translated.
Two dub versions reference the same wind, traffic and score objects but different dialogue objects. For the acceptance record, identify which audioProgramme and component IDs are involved and what the receiver should hear. Record the relevant language choices and graph links so that a second engineer can reproduce the test rather than trusting an assertion that the XML was translated.
Duplicating shared audio or changing its gain differently by accident can produce inconsistent programmes that are not merely language variants. This defect can remain undetected in a text-only inspection because the metadata may still parse and the media may still open. The real test must check both the intended model and its audible consequences.
Compare the common content references and playback of shared stems across language selections. Record source and target values, the applicable standard or production profile, the rendering tool and any exception. Use the same acceptance conditions for the original and localized versions so differences can be explained precisely.
Default programme selection
A playback system may need a default programme when no explicit user choice is provided. Delivery policy must account for that behavior. A localized audio programme can be grammatically labeled correctly and still be the wrong programme. Its language metadata, linked content, object relationships and rendering behavior must agree. The appropriate review therefore follows the user’s choice all the way to audible output.
A device starts playback without a preferred language. The metadata graph contains several valid programme alternatives. A side-by-side comparison should distinguish allowed display-text changes from invariant links, timing and sample mappings. This prevents the team from confusing a metadata correction with a new audio mix. When a different mix is needed, explicitly open a mastering task with its own evidence.
An unreviewed default could play the wrong language or an accessibility version when a distributor expects a specific base presentation. This defect can remain undetected in a text-only inspection because the metadata may still parse and the media may still open. The real test must check both the intended model and its audible consequences.
Agree on the intended default behavior, test without a language preference and document the actual selection made by the target renderer. Record source and target values, the applicable standard or production profile, the rendering tool and any exception. Use the same acceptance conditions for the original and localized versions so differences can be explained precisely.
Language naming and accessibility labels
Human-readable programme and object names can be localized to aid navigation, but identifiers and language codes remain structured facts. The broad ADM model is useful because it separates what a programme means from how its audio is structured. That flexibility demands care: a translated value may be descriptive, a code may be structural and a numeric field may affect the listener’s spatial or temporal experience.
A receiver displays an audio menu with English, Español and a descriptive-audio option; labels must help the listener choose correctly. The receiving system’s behavior is authoritative for interoperability. A generic XML reader can confirm syntax but cannot demonstrate that a playout pipeline understands complementary objects, gain limits or BW64 channel allocation. Test with the intended renderer and document limitations.
Calling audio description simply English may hide an accessibility feature; using an unsupported language-tag syntax may break automatic selection. This defect can remain undetected in a text-only inspection because the metadata may still parse and the media may still open. The real test must check both the intended model and its audible consequences.
Review menu labels, programme roles and language tags together, including how screen readers announce the options. Record source and target values, the applicable standard or production profile, the rendering tool and any exception. Use the same acceptance conditions for the original and localized versions so differences can be explained precisely.
Protect rendering instructions and technical meaning
Audio pack and channel formats
audioPackFormat and audioChannelFormat help describe channel groupings, speaker feeds and other format-dependent behavior. Multilingual delivery is a design question as well as a language question. Decide whether the recipient wants independent fixed presentations, interactive alternatives or separate exported masters. Each path has different metadata and testing needs. A translator should not silently choose an interaction model on behalf of a broadcaster.
A 5.1 bed and a mono dialogue object require different format descriptions even if both are part of the same programme. Where there is disagreement, investigate it at the correct layer: language label, language code, graph reference, channel mapping, sample content or renderer configuration. Altering a translated title cannot repair a missing track. A structured defect report shortens the route to a genuine fix.
Renaming an audio pack must not change type definitions or which channels it contains. Misclassified formats can steer audio to wrong speakers. This defect can remain undetected in a text-only inspection because the metadata may still parse and the media may still open. The real test must check both the intended model and its audible consequences.
Compare pack references, channel-format relationships and output-channel behavior on the intended rendering setup. Record source and target values, the applicable standard or production profile, the rendering tool and any exception. Use the same acceptance conditions for the original and localized versions so differences can be explained precisely.
Objects, positions and movement
Object-based audio can carry spatial information that affects rendering. Position values are technical data rather than directions to translate. The most reliable proof is reproducible playback. Editors can compare titles, identifiers and language tags, but only an actual render can show the correct narrator, shared ambience, spatial motion and accessible selection. Treat both the data graph and the rendered experience as deliverables.
A helicopter sound moves across the listener’s sound field while dialogue remains anchored. The localized narrator does not alter that trajectory. Maintain a small representative test suite with at least two languages, shared audio, a spatial event and a known default selection. Run it before and after localization. It is easier to catch a graph regression with a focused example than at the end of a long programme.
Translating left or right as if they were free-text position fields, or modifying numeric coordinates, can move the audible scene. This defect can remain undetected in a text-only inspection because the metadata may still parse and the media may still open. The real test must check both the intended model and its audible consequences.
Compare object position data and audition the resulting motion using a compliant renderer before accepting changes. Record source and target values, the applicable standard or production profile, the rendering tool and any exception. Use the same acceptance conditions for the original and localized versions so differences can be explained precisely.
Start, duration and timing alignment
Objects and programmes can have start and duration information. Language versions may differ in speech timing but must meet agreed synchronization constraints. Think in paths through the audio graph. The listener does not hear a text field; the listener hears the samples selected and rendered through a chain of references. Names help operators understand those paths, but they cannot substitute for the technical relationships. This distinction governs which properties may be translated.
A localized narration has an approved new audio track that still enters at the same scene boundary as the original programme. For the acceptance record, identify which audioProgramme and component IDs are involved and what the receiver should hear. Record the relevant language choices and graph links so that a second engineer can reproduce the test rather than trusting an assertion that the XML was translated.
Changing metadata timing to conceal a poorly aligned dub can alter object availability or introduce gaps at programme transitions. This defect can remain undetected in a text-only inspection because the metadata may still parse and the media may still open. The real test must check both the intended model and its audible consequences.
Check timing against picture reference, object activation and duration, then document any editorially approved timing adaptation. Record source and target values, the applicable standard or production profile, the rendering tool and any exception. Use the same acceptance conditions for the original and localized versions so differences can be explained precisely.
Gain, importance and interaction
ADM can describe parameters that influence rendering and user control, including gain and object interaction constraints. A localized audio programme can be grammatically labeled correctly and still be the wrong programme. Its language metadata, linked content, object relationships and rendering behavior must agree. The appropriate review therefore follows the user’s choice all the way to audible output.
A user may be allowed to change commentary level without changing the fixed ambience. The limits are part of the designed experience. A side-by-side comparison should distinguish allowed display-text changes from invariant links, timing and sample mappings. This prevents the team from confusing a metadata correction with a new audio mix. When a different mix is needed, explicitly open a mastering task with its own evidence.
A translated label that suggests an object is adjustable when it is not can mislead users, while changing actual gain ranges affects playback. This defect can remain undetected in a text-only inspection because the metadata may still parse and the media may still open. The real test must check both the intended model and its audible consequences.
Inspect interactive controls, permitted ranges, importance values and audible gain on the recipient’s playback platform. Record source and target values, the applicable standard or production profile, the rendering tool and any exception. Use the same acceptance conditions for the original and localized versions so differences can be explained precisely.
Loudness metadata per content and programme
Loudness metadata communicates measured audio behavior. It must be associated with the correct component or presentation. The broad ADM model is useful because it separates what a programme means from how its audio is structured. That flexibility demands care: a translated value may be descriptive, a code may be structural and a numeric field may affect the listener’s spatial or temporal experience.
A French programme with a different narrator may have different measured integrated loudness from the English version. The receiving system’s behavior is authoritative for interoperability. A generic XML reader can confirm syntax but cannot demonstrate that a playout pipeline understands complementary objects, gain limits or BW64 channel allocation. Test with the intended renderer and document limitations.
Copying the English loudness values into a new mix without measurement is inaccurate. A language change can alter the programme’s overall signal. This defect can remain undetected in a text-only inspection because the metadata may still parse and the media may still open. The real test must check both the intended model and its audible consequences.
Measure the actual version under the delivery policy and check that programme and content metadata identify the right results. Record source and target values, the applicable standard or production profile, the rendering tool and any exception. Use the same acceptance conditions for the original and localized versions so differences can be explained precisely.
Downmixes and rendering variations
The same ADM scene may be rendered for different speaker layouts or devices. Descriptive translation must not mask audible differences. Multilingual delivery is a design question as well as a language question. Decide whether the recipient wants independent fixed presentations, interactive alternatives or separate exported masters. Each path has different metadata and testing needs. A translator should not silently choose an interaction model on behalf of a broadcaster.
A living-room TV and headphones render the same programme differently, and users still need clearly labeled dialogue selections. Where there is disagreement, investigate it at the correct layer: language label, language code, graph reference, channel mapping, sample content or renderer configuration. Altering a translated title cannot repair a missing track. A structured defect report shortens the route to a genuine fix.
A mix that seems correct on one monitoring setup may hide a missing object or bad speaker assignment in another. This defect can remain undetected in a text-only inspection because the metadata may still parse and the media may still open. The real test must check both the intended model and its audible consequences.
Validate representative target layouts and device classes, including stereo fallback where required by the delivery agreement. Record source and target values, the applicable standard or production profile, the rendering tool and any exception. Use the same acceptance conditions for the original and localized versions so differences can be explained precisely.
Test the multilingual experience end to end
Schema and production-profile validation
The general ADM standard is broad; broadcast profiles can narrow it for reliable interchange. The recipient decides which constraints apply. The most reliable proof is reproducible playback. Editors can compare titles, identifiers and language tags, but only an actual render can show the correct narrator, shared ambience, spatial motion and accessible selection. Treat both the data graph and the rendered experience as deliverables.
A facility supports the EBU ADM production profile and declines a graph that uses a less common feature without an approved conversion. Maintain a small representative test suite with at least two languages, shared audio, a spatial event and a known default selection. Run it before and after localization. It is easier to catch a graph regression with a focused example than at the end of a long programme.
A well-formed XML file may fail profile validation, and renaming fields will not remedy unsupported graph structure. This defect can remain undetected in a text-only inspection because the metadata may still parse and the media may still open. The real test must check both the intended model and its audible consequences.
Validate against the applicable ITU recommendation and agreed production profile, including newer EBU guidance where required. Record source and target values, the applicable standard or production profile, the rendering tool and any exception. Use the same acceptance conditions for the original and localized versions so differences can be explained precisely.
Reference graph integrity
XML parsing cannot by itself establish that all references form a coherent ADM scene with available audio. Think in paths through the audio graph. The listener does not hear a text field; the listener hears the samples selected and rendered through a chain of references. Names help operators understand those paths, but they cannot substitute for the technical relationships. This distinction governs which properties may be translated.
A programme links to content that links to an object that in turn points to a missing track UID. For the acceptance record, identify which audioProgramme and component IDs are involved and what the receiver should hear. Record the relevant language choices and graph links so that a second engineer can reproduce the test rather than trusting an assertion that the XML was translated.
The file can appear syntactically normal even though the selected presentation contains silence or the wrong audio component. This defect can remain undetected in a text-only inspection because the metadata may still parse and the media may still open. The real test must check both the intended model and its audible consequences.
Walk the graph from each audioProgramme to its audioContent, audioObjects, track UIDs and formats; fail on unintended missing edges. Record source and target values, the applicable standard or production profile, the rendering tool and any exception. Use the same acceptance conditions for the original and localized versions so differences can be explained precisely.
Language-selection playback tests
The same file can offer several choices; each choice is a separate user-visible result requiring review. A localized audio programme can be grammatically labeled correctly and still be the wrong programme. Its language metadata, linked content, object relationships and rendering behavior must agree. The appropriate review therefore follows the user’s choice all the way to audible output.
A listener selects French but hears original English narration mixed underneath because an old object remained referenced. A side-by-side comparison should distinguish allowed display-text changes from invariant links, timing and sample mappings. This prevents the team from confusing a metadata correction with a new audio mix. When a different mix is needed, explicitly open a mastering task with its own evidence.
A text-only QA report showing two language labels cannot detect simultaneous voices or missing language-independent beds. This defect can remain undetected in a text-only inspection because the metadata may still parse and the media may still open. The real test must check both the intended model and its audible consequences.
Audition each supported selection and confirm the menu label, audible dialogue, shared effects and accessibility behavior. Record source and target values, the applicable standard or production profile, the rendering tool and any exception. Use the same acceptance conditions for the original and localized versions so differences can be explained precisely.
Interaction and accessibility testing
Interactive objects and accessible presentations must behave as described, not merely appear as menu entries. The broad ADM model is useful because it separates what a programme means from how its audio is structured. That flexibility demands care: a translated value may be descriptive, a code may be structural and a numeric field may affect the listener’s spatial or temporal experience.
An audio description programme and an alternate commentary channel are offered alongside standard narration. The receiving system’s behavior is authoritative for interoperability. A generic XML reader can confirm syntax but cannot demonstrate that a playout pipeline understands complementary objects, gain limits or BW64 channel allocation. Test with the intended renderer and document limitations.
If the wrong control is exposed, listeners may lose essential description or inadvertently mute important speech. This defect can remain undetected in a text-only inspection because the metadata may still parse and the media may still open. The real test must check both the intended model and its audible consequences.
Test the actual interface and renderer controls, verify interaction limits and ask representative users to review accessibility labels. Record source and target values, the applicable standard or production profile, the rendering tool and any exception. Use the same acceptance conditions for the original and localized versions so differences can be explained precisely.
Version, profile and tool interoperability
Different ingest, DAW and rendering applications may support different versions or subsets of ADM and BW64. Multilingual delivery is a design question as well as a language question. Decide whether the recipient wants independent fixed presentations, interactive alternatives or separate exported masters. Each path has different metadata and testing needs. A translator should not silently choose an interaction model on behalf of a broadcaster.
An international production passes the same file among a mixing studio, quality-control vendor and playout platform. Where there is disagreement, investigate it at the correct layer: language label, language code, graph reference, channel mapping, sample content or renderer configuration. Altering a translated title cannot repair a missing track. A structured defect report shortens the route to a genuine fix.
A tool can silently omit metadata elements it does not understand while keeping the audio samples intact. This defect can remain undetected in a text-only inspection because the metadata may still parse and the media may still open. The real test must check both the intended model and its audible consequences.
Run a round-trip with the target tools, compare graph elements and note unsupported semantics and approved conversions. Record source and target values, the applicable standard or production profile, the rendering tool and any exception. Use the same acceptance conditions for the original and localized versions so differences can be explained precisely.
Acceptance, provenance and repair path
A complete handoff identifies source files, media identity, approved language choices, profile constraints and verification evidence. The most reliable proof is reproducible playback. Editors can compare titles, identifiers and language tags, but only an actual render can show the correct narrator, shared ambience, spatial motion and accessible selection. Treat both the data graph and the rendered experience as deliverables.
A studio issues a new Portuguese version and must prove its programme selection is correct while keeping the English source available for rollback. Maintain a small representative test suite with at least two languages, shared audio, a spatial event and a known default selection. Run it before and after localization. It is easier to catch a graph regression with a focused example than at the end of a long programme.
Without a version register, a recipient may confuse a metadata-only correction with a new dub or unknowingly play a previous mix. This defect can remain undetected in a text-only inspection because the metadata may still parse and the media may still open. The real test must check both the intended model and its audible consequences.
Keep immutable sources, revisions, test results, renderer version, audible spot checks and a clear approval record for every delivery. Record source and target values, the applicable standard or production profile, the rendering tool and any exception. Use the same acceptance conditions for the original and localized versions so differences can be explained precisely.
Three realistic international delivery scenarios
Two narrations with common music and ambience
Create an inventory of the content graph before proposing localized changes. The English and Spanish presentations must each select the appropriate dialogue component while retaining the approved shared ambience and music. Preserve existing IDs for unchanged elements. If a new audio object is added, treat it as an authorized mastering operation with corresponding valid IDs, mappings and media. Render and listen to both programmes. A translated menu label alone does not create a new language version.
The takeaway is that the language variant is a complete listener-facing selection, not simply translated words on an engineering screen. Define the intended choice, preserve graph identity and sample mapping, render the selected audio and record that the result conforms to the brief. The receiving team needs evidence it can reproduce.
Interactive commentary switching during live sport
First determine whether the listener actually controls language or commentary selection, rather than merely choosing a programme before playout. Validate complementary object relationships and the interaction policy. Ensure that shared stadium sound remains available and that selecting one commentary does not accidentally enable two languages simultaneously. Test controls under playback conditions that resemble the recipient’s device. User experience is part of technical acceptance, not an optional cosmetic extra.
The takeaway is that the language variant is a complete listener-facing selection, not simply translated words on an engineering screen. Define the intended choice, preserve graph identity and sample mapping, render the selected audio and record that the result conforms to the brief. The receiving team needs evidence it can reproduce.
An archival BW64 file sent to another country
Use the specified ADM and BW64 versions and confirm that the partner’s parser understands the metadata. Validate axml, chna and every track UID connection; keep the original carrier file immutable. Localize only approved titles and descriptions, and keep language codes and numeric spatial data correctly typed. Produce a recipient test render and record any features lost in conversion. A long-form file is not a reliable multilingual archive unless its metadata survives the real transfer.
The takeaway is that the language variant is a complete listener-facing selection, not simply translated words on an engineering screen. Define the intended choice, preserve graph identity and sample mapping, render the selected audio and record that the result conforms to the brief. The receiving team needs evidence it can reproduce.
Frequently asked questions
Is ADM the same thing as a codec?
No. ADM describes audio scene and content relationships. It can accompany audio carried in a compatible file or workflow, and the chosen coding and rendering systems still matter.
Will translating audioProgrammeName change the spoken language?
No. The referenced content and tracks determine the audio. A label change without the appropriate audio selection produces a misleading programme.
Why do axml and chna both matter?
axml carries the XML description of the scene, while chna links sample-bearing file tracks to ADM identifiers. Both relationships must agree in a BW64 interchange workflow.
Can music and ambience be shared across language versions?
Yes, where the intended programmes reference common components. Verify that each version still selects the right language-dependent dialogue.
How do interactive language choices differ from separate programmes?
They use different modeling and playback behavior. Choose the approach permitted by the implementation and delivery brief and test the actual renderer.
What makes a localized ADM file ready for delivery?
Valid format and profile, complete graph references, correct channel allocation, actual audio in every offered language, acceptable rendering and recorded reviewer approval.
Final technical and linguistic acceptance checklist
Verify agreed ITU and EBU document versions, the chosen production profile, complete audioProgramme and audioContent relationships, audioObject identity, track UID resolution, axml and chna consistency, BW64 carrier integrity, available dialogue languages, shared content, default selection, spatial information, interaction constraints, loudness, timing and user-facing labels. Render every offered language version with the intended recipient tools, listen critically and retain immutable source media and a documented sign-off. A good ADM localization is one that people can both understand and hear correctly.
Authoritative specifications and guidance
ITU-R BS.2076-3: Audio Definition Model (2025); ITU-R BS.2088: BW64 long-form audio format; EBU Tech 3393: ADM production profile; EBU ADM: BW64, axml and chna; EBU ADM: multiple-language use cases.
Related eduKateSG guides
For the neighboring but different tasks, see BWF bext and location-sound metadata, EBUCore descriptive media metadata and IMF audiovisual mastering versions. ADM addresses the actual composition and rendering of a structured audio scene, rather than the whole media catalogue or the interchange of a video timeline.