Languages - AgentTV

Every video has a spoken language, and it can have subtitle tracks in other languages. AgentTV records both for every account and every source. AI apps and search engines read them, and answers use them to tell the asker what language a clip is in. All language values are BCP 47 tags (the standard web language codes), for example es, de-AT or en-US. When AgentTV does not know a language, it leaves the field out. It never guesses from the title, the account or the country, and it never defaults to English.

Spoken language

The spoken language is the language people speak in the video. AgentTV sets it from the first of these that exists:

Owner. A correction for one video. Tell us when a detected language is wrong, and we set it.
Detected. The transcription finds it in the audio.
Feed. Your catalog feed declares it. This value holds only until a transcript exists.

language_source says which of the three set it (owner, detected or feed). language_confidence is the detection confidence from 0 to 1, and is set only for a detected language. Transcripts, chapter quotes and moment text stay in the spoken language. AgentTV does not translate them.

Subtitle tracks

A subtitle track is a file your source provides, such as a caption track your metadata API lists or an HLS subtitle rendition. Each track has a language, a kind (human or auto) and a url. Our own transcript is not a subtitle track. caption_languages lists the languages of your human tracks. This is the list answers and labels name. A track URL appears in public output only when it is a WebVTT file (.vtt) at a public https address with no signed or expiring token, and only when your rights allow quoting the video’s words. An HLS subtitle playlist (.m3u8) is a stream address, not a caption file, so it is never published. In every other case the track keeps its language and its url is null.

The language fields

The same fields appear on the page block JSON and in the ARV record:

{
  "language_primary": "es",
  "language_source": "detected",
  "language_confidence": 0.97,
  "subtitle_tracks": [
    { "language": "en", "kind": "human", "url": "https://cdn.example.com/subs/final.en.vtt" }
  ],
  "caption_languages": ["en"]
}

Search results, moments and citations from the API and MCP carry the short form: language_primary and caption_languages.

Where the language appears

Surface	What it carries
JSON-LD `VideoObject` (watch page and page block)	`inLanguage`, `subtitleLanguage` and `caption`
Page block JSON	the language fields above
Page block `body_html`	a `lang` attribute on the transcript and chapter text
ARV record	the asset `language` and `caption_languages`, plus the full fields as an extension
Search, moments and citations (API and MCP)	`language_primary` and `caption_languages`
JSON Feed (`feed.json`)	the item `language`
`llms.txt` and the watch page Markdown	the spoken language and the subtitle languages

JSON-LD

The page block and the watch page follow Google’s VideoObject guidance and schema.org:

{
  "@type": "VideoObject",
  "inLanguage": "es",
  "subtitleLanguage": ["en", "de"],
  "caption": [
    { "@type": "MediaObject", "inLanguage": "en", "encodingFormat": "text/vtt", "contentUrl": "https://cdn.example.com/subs/final.en.vtt" },
    { "@type": "MediaObject", "inLanguage": "de", "encodingFormat": "text/vtt", "contentUrl": "https://cdn.example.com/subs/final.de.vtt" }
  ]
}

Clip items under hasPart inherit the language from the VideoObject, so they do not repeat it. The same rights rules apply as to the rest of the markup: a video that may not be cited gets no markup at all.

ARV

The ARV 1.0 asset states the spoken language as language and the human subtitle languages as caption_languages. The full language fields use the standard’s extension mechanism, under the com.agentcdn key:

{
  "extensions": {
    "com.agentcdn": {
      "video_language": {
        "language_primary": "es",
        "language_source": "detected",
        "language_confidence": 0.97,
        "subtitle_tracks": [
          { "language": "en", "kind": "human", "url": "https://cdn.example.com/subs/final.en.vtt" }
        ],
        "caption_languages": ["en"]
      }
    }
  }
}

ARV 1.0 has no moment language field, so each moment carries extensions["com.agentcdn"].language_primary and caption_languages, inherited from its video.

Answers in the asker’s language

Ask and the AI app tools answer in the language of the question, whatever language the videos are spoken in. The answer plan states that language as answer_plan.language. When a cited clip is spoken in another language, the answer says so once. It names subtitles in the asker’s language only when the video has a human track in that language. For example: “This clip is in Spanish, with English subtitles.” A clip in the asker’s language, or in no known language, is never named for its language.

​Spoken language

​Subtitle tracks

​The language fields

​Where the language appears

​JSON-LD

​ARV

​Answers in the asker’s language