OCP Pipeline

The OCP (OVOS Common Playback) framework integrates seamlessly into ovos-core to enhance media playback through an intent-aware, classifier-assisted pipeline. It enables automatic recognition and handling of media-related queries, such as "play music", "pause video", or "next song", using trained classifiers, skill catalogs, and playback-specific filters.

Pipeline Components

The OCP Pipeline includes multiple components, designed to detect and handle media-related intents with varying levels of certainty. These components are placed in order of confidence in the intent pipeline configuration:

1. ocp_high - Unambiguous

  • Purpose: Catches queries that are clearly and explicitly about media playback.
  • Behavior: Triggers only when the user's request unambiguously indicates a media action, e.g., "play Metallica".
  • Usage: Should be placed before other components in the pipeline.
{
  "intents": {
    "pipeline": [
      "converse",
      "ocp_high",
      "fallback_low"
    ]
  }
}

2. ocp_medium - Semi-Ambiguous

  • Purpose: Handles likely media queries where intent is probable but not definite.
  • Behavior: Uses a binary classifier to determine if the utterance is media-related.
  • Usage: Placed after ocp_high, to catch softer matches.
{
  "intents": {
    "pipeline": [
      "converse",
      "ocp_medium",
      "fallback_low"
    ]
  }
}

3. ocp_low - Ambiguous

  • Purpose: Captures uncertain queries using keyword-based heuristics.
  • Behavior: Triggers whenever the query includes a known media-related keyword (e.g., artist, movie name) registered by a skill.
  • Usage: Last media filter, before true fallback layers.
{
  "intents": {
    "pipeline": [
      "converse",
      "ocp_low",
      "fallback_low"
    ]
  }
}

Skill Media Keywords

OCP-aware skills can register domain-specific media keywords — such as artist names, albums, show titles, etc. — that enhance intent recognition and classification.

These keywords serve multiple purposes:

  • Improve query classification (e.g., "play Enter Sandman" → MUSIC).
  • Influence classifier bias when media type is uncertain (e.g., "Matrix" might bias toward MUSIC if it's registered only as a song).
  • Expand the available media catalog dynamically as new skills are added.

⚠️ The ocp_low component relies on these keywords and will initiate a media search whenever known media is present on the text, even on questions unrelated to media playback

MediaType Matching

To improve accuracy and reduce search latency, OCP classifies user queries into media types:

Examples of supported media types:

  • music
  • movie
  • radio
  • podcast
  • audiobook
  • news

Media type is inferred in two ways:

  • Keyword-Based Classification (default): If the query includes words like "movie", "news", or "podcast", OCP assigns the relevant media type.
  • Classifier-Based Matching (optional/experimental): An experimental classifier (currently English-only) predicts the media type based on the full query and context. This classifier adapts its bias dynamically using registered skills as a catalog — for instance, if a skill registers "Matrix" as a song, the classifier is more likely to infer MUSIC rather than MOVIE.

The default implementation is currently quite limited, each keyword triggers its corresponding MediaType. For example:

  • A query containing the keyword "movie" will be categorized as MediaType.MOVIE.
  • A query with "news" will be classified as MediaType.NEWS.
  • A query with "radio" will be categorized under MediaType.RADIO.

💡 A query is initially classified as MediaType.GENERIC unless it explicitly contains a keyword

Result Filtering

After potential media results are gathered from OCP-enabled skills, the pipeline filters them based on a series of rules to ensure only valid and playable entries are passed to the playback engine.

  • Confidence Score Threshold: Each result includes a match_confidence score (0–100). If it falls below the configured min_score threshold, it is discarded.
  • Media Type Filtering: If the query was classified as a specific media type (e.g., MUSIC), results with a different media type (e.g., PODCAST) are removed.
  • Plugin Availability (Stream Extractors): Each result has a URI scheme (e.g., http://, spotify://). If the required plugin is not installed (like ovos-media-spotify), the result is discarded to prevent playback errors.
  • Audio/Video Preference Filtering: If the user (or system) requests audio-only or video-only playback
  • Audio mode skips video-only results.
  • Video mode skips pure audio unless a video is attached.
  • The pipeline dynamically adapts if the query includes hints like "watch" or "just audio".

Playback Frameworks

Once the best match is selected, playback is delegated to the appropriate media plugin loaded via ovos-audio.

Skills do not handle media playback directly — they return media candidates, and OCP takes care of launching and controlling the stream.

💡 Skills act as media catalogs. Playback is handled by ovos-audio, and only skills with compatible stream URIs are selected.

OCP Configuration Options

All of the above behavior is customizable using config settings under "OCP" in the "intents" subsection of your mycroft.conf.

{
"intents": {
    "OCP": {
      "experimental_media_classifier": false,
      "experimental_binary_classifier": false,
      "legacy": false,
      "classifier_threshold": 0.4,
      "min_score": 40,
      "filter_media": true,
      "filter_SEI": true,
      "playback_mode": 0,
      "search_fallback": true
    }
}
}
Key Type Default Description
experimental_media_classifier bool false Enable ML-based media type classification (English only).
experimental_binary_classifier bool false Enable ML-based media detection for ocp_medium (is this query media-related?).
legacy bool false Use legacy audio service bus api instead of OCP bus api (not recommended).
classifier_threshold float 0.4 Minimum confidence for trusting classifier result (0.0–1.0).
min_score int 40 Minimum match confidence for keeping a skill result (0–100).
filter_media bool true Enable media type-based result filtering.
filter_SEI bool true Filter out results that require unavailable plugins (Stream Extractors).
playback_mode int 0 Playback preference: 0 = auto, 10 = audio-only, 20 = video-only.
search_fallback bool true perform a second GENERIC media seach if no type-specific results are found.

Summary

  • OCP uses classifiers, keyword matching, and skill-registered catalogs to intelligently route and filter media playback requests.
  • It prioritizes clarity of intent through configurable components (ocp_high, ocp_medium, ocp_low).
  • Post-matching filters guarantee high confidence, media type compatibility, plugin support, and preferred output mode.
  • Playback is entirely managed by ovos-audio using the selected plugin backend — skills never handle actual playback.

Media Query Embeddings (upcoming)

image

image