archy/apps/pine-whisper/manifest.yml

app:
  id: pine-whisper
  name: Pine Whisper (STT)
  # App revision 3.4.2 = upstream wyoming-whisper 3.4.1 image + tuned args
  # (--beam-size 1). Bumped past the image version so catalog-driven nodes
  # pick up the args change; the pre-release form "3.4.1-1" would compare
  # LOWER than 3.4.1 under semver and never roll out.
  version: "3.4.2"
  # Tracks the rhasspy/wyoming-whisper image we pin (Docker Hub — the
  # project's GitHub tags are not the image tags). NOTE: this manifest
  # deliberately ships an args-tuned revision AHEAD of the image tag (see
  # comment above) — BEHIND here means the image tag moved and the tuned
  # revision needs re-basing onto it, not just a pin bump.
  upstream:
    kind: dockerhub
    repo: rhasspy/wyoming-whisper
  description: Wyoming-protocol faster-whisper speech-to-text engine. Internal Pine voice-assistant stack member — turns speech captured by a PineVoice satellite into text for Home Assistant Assist.
  category: home

  # Hyphen name matches the runtime references (stack member table / startup
  # order) + the live container, so on an existing node the orchestrator ADOPTS
  # the running engine rather than recreating it (downloaded models under /data
  # preserved).
  container_name: pine-whisper

  container:
    image: docker.io/rhasspy/wyoming-whisper:3.4.1
    pull_policy: if-not-present
    network: archy-net
    network_aliases: [pine-whisper]
    # The image entrypoint already binds tcp://0.0.0.0:10300; these args only
    # pick the model + language (mirrors the pine ha-stack.yml compose command).
    # --beam-size 1: the image default is 5 on x86 (1 on ARM). Benchmarked on
    # a test node (i5-1135G7, base-int8): beam 1 transcribes the same text
    # ~45% faster — the standard low-latency setting for short voice commands
    # (HA's own whisper add-on defaults to 1).
    custom_args: ["--model", "base-int8", "--language", "en", "--beam-size", "1"]

  dependencies:
    - storage: 2Gi

  resources:
    memory_limit: 2Gi

  security:
    # cap-drop=ALL is applied by the orchestrator. A plain Python Wyoming server
    # on an unprivileged port needs no added capabilities.
    capabilities: []
    readonly_root: false   # downloads the whisper model into /data on first run
    no_new_privileges: true
    network_policy: isolated

  ports:
    # Published so Home Assistant (on the pasta net) can reach the engine via
    # host.containers.internal:10300 (the Wyoming integration endpoint).
    - host: 10300
      container: 10300
      protocol: tcp
      auth: none
      auth_rationale: >-
        Wyoming voice protocol, a binary local-only stream consumed by Home Assistant; not HTTP and not browser-reachable.

  volumes:
    - type: bind
      source: /var/lib/archipelago/pine-whisper
      target: /data
      options: [rw]

  environment: []

  health_check:
    type: tcp
    endpoint: localhost:10300
    interval: 30s
    timeout: 5s
    retries: 5
    start_period: 60s   # first start downloads the model

  metadata:
    author: Rhasspy / Home Assistant
    icon: /assets/img/app-icons/pine.svg
    website: https://github.com/rhasspy/wyoming-faster-whisper
    repo: https://github.com/rhasspy/wyoming-faster-whisper
    license: MIT
    tags:
      - home
      - voice
      - speech-to-text
      - wyoming