Taxonomy

This page covers the leaf axis, mediavocab.MediaType, and the genre tags that ride alongside it. It is one of four orthogonal axes; the coarse axes (domain, modality, structure) and the rationale for splitting them are in classification-model.md.

mediavocab is the source of truth

This package does not define its own media taxonomy. The public output type is mediavocab.MediaType, a string-Enum owned by the mediavocab package, and it is re-exported here as MediaType purely for convenience:

from ovos_media_classifier import MediaType   # this IS mediavocab.MediaType
from mediavocab import MediaType as MV
assert MediaType is MV

The classifier's classify() return value is mapped onto a mediavocab.MediaType at the boundary, so the public contract enforces the shared vocabulary.

Type + genres, not a per-type intent

The classifier models a play request along the real axes only: the OCPDomain (play / control / not-ocp) × a mediavocab.MediaType + mediavocab genre tags. There is no separate per-media-type intent enum. Distinctions the public taxonomy treats as genre or content-form rather than as media types, anime, cartoon, asmr, and the adult variants, are not their own MediaType: they resolve to a base type and carry their nuance as a genre tag.

Backends emit a raw detection label (a .voc / model-head string such as "music", "adult", "anime") which resolves directly to (MediaType, genres) via two string-keyed maps in ovos_media_classifier/intents.py:

• LABEL_TO_MEDIA_TYPE: dict[str, MediaType], the base type.
• LABEL_TO_GENRES: dict[str, list[str]], genre tags preserving the nuance the type drops. Only tags present in mediavocab's KNOWN_GENRES are emitted, so genres are taxonomy-enforced too. The helper genres_for_label(label) -> list[str] reads this map.

Representative examples:

raw label	→ MediaType	→ genres
music	MUSIC	,
movie	MOVIE	,
tv_show	EPISODIC_SERIES	,
anime	EPISODIC_SERIES	["anime"]
cartoon	EPISODIC_SERIES	["animation"]
asmr	PROCEDURAL_AMBIENT	["asmr"]
ambient	PROCEDURAL_AMBIENT	,
audiobook	AUDIOBOOK	,
book	BOOK	,
comic	COMIC	,
playlist	PLAYLIST	,
sound_effect	SOUND_EFFECT	,
interactive_fiction	INTERACTIVE_FICTION	,
adult	MOVIE	["adult"]
adult_audio	MUSIC	["adult"]
hentai	EPISODIC_SERIES	["anime", "adult"]
news	RADIO	,
documentary	MOVIE	,

So a query like "play some hentai" yields a MediaType.EPISODIC_SERIES from classify() and ["anime", "adult"] from classify_genres(). The adult tag is what the content filter blocks on, the type alone never carries that signal.

Every non-sentinel mediavocab.MediaType is reachable from a raw label, so the whole taxonomy is exercised. A few distinctions are carried by how a work is consumed: audiobook (play a narration → AUDIOBOOK) vs book (TTS-read a text → BOOK); anime (watch → EPISODIC_SERIES) vs comic/manga (read → COMIC); asmr/ambient (both PROCEDURAL_AMBIENT, the former genre-tagged).

Domains and control intents

Above media types sits OCPDomain, the top-level question "does this command target OCP at all?":

OCPDomain	Meaning
OCP_PLAY	a playback request, the what is a MediaType + genres
OCP_CONTROL	a player-control request, refine with OCPControlIntent
NOT_OCP	unrelated to media playback

OCPControlIntent enumerates control actions (PLAY, PAUSE, STOP, NEXT, SHUFFLE, SEEK_FORWARD, …). Only a classifier with a dedicated control head classifies these; the bundled keyword classifier derives the domain from the media-type result and so reports OCP_PLAY / NOT_OCP only. A backend that models control intents detects OCP_CONTROL.

Entity labels

OCPEntityLabel is the NER vocabulary (artist_name, movie_title, tv_show_title, radio_station, …). Each entity label maps directly to a mediavocab.MediaType via NER_LABEL_TO_MEDIA_TYPE (and to genre tags via NER_LABEL_TO_GENRES), so an entity hit ("play Dune: Part Two") resolves to a media type, and an adult_title / pornstar hit still surfaces the adult genre the content filter blocks on.

The NER backend (AhocorasickMediaClassifier) uses these labels to tag the user's entity lists. They are also exported so that an out-of-tree classifier, and the skills that register content under these labels at runtime, shares the same vocabulary.

Query vs. content classification

This package classifies a command, the natural-language request a user speaks. That is a different problem from classifying a piece of content:

	This package	mediavocab.text
Input	a voice command, e.g. "play the news"	catalog metadata, e.g. a video title + description
Question	what does the user want OVOS to do?	what kind of thing is this item?
Output	(MediaType, conf) + domain + genres	ClassificationResult for the item
Used by	the OCP pipeline (intent stage)	catalog tagging / ingestion

They share the mediavocab.MediaType vocabulary but answer opposite questions: intent-of-a-request here, type-of-an-item there. Do not substitute one for the other.

---

← Extending · Home · Backends →