Skip to content

Agent Zero AI framework

2.6 KB
Raw Blame

Kokoro TTS Plugin DOX

Purpose

  • Own local Paradee/Kokoro text-to-speech integration and browser TTS fallback coordination under the stable _kokoro_tts plugin identity.

Ownership

  • api/ owns synthesize and status endpoints.
  • helpers/runtime.py owns voice normalization and model dispatch; helpers/paradee.py owns the CPU ONNX runtime and phonemizer.
  • helpers/migration.py preserves the legacy tts_kokoro enable/disable setting.
  • hooks.py owns config hooks and idempotent Paradee dependency setup. Plugin-owned startup and enable extensions invoke the same ensure_dependencies hook.
  • webui/ owns settings and speech UI integration.
  • default_config.yaml, plugin.yaml, and README.md own defaults, metadata, and behavior notes.

Local Contracts

  • Additional Paradee dependencies (ONNX Runtime and the English spaCy model) belong to hooks.py, targeted at the current framework interpreter. Setup runs on install/update, enabled-plugin startup, and re-enable; already-current environments skip installation. Failures are logged and leave browser fallback available. Never install during synthesis or config reads.
  • Preserve browser-native fallback when the plugin is disabled or unavailable.
  • Do not expose generated speech artifacts outside intended response paths.
  • Preserve voice and speed configuration compatibility; weighted blends stay in memory and use positive finite weights.
  • voice: paradee selects the single American English Heart voice and is the default for unconfigured settings. Saved Kokoro voices, blends, speed, and enable/disable choices remain unchanged; no provider-choice migration or dialog is needed.
  • Valid Kokoro weights take precedence over the voice string. Paradee is never a weighted voice; mixed expressions discard its reserved token and retain the Kokoro selection.
  • Paradee model assets stay under ignored usr/plugins/_kokoro_tts/paradee/, pinned to the Hugging Face v1.0 revision. Long text is split without dropping phonemes and generated audio remains in response memory.
  • Model preload and status must describe the selected engine. The frontend voice picker replaces the selection when crossing between Paradee and Kokoro.

Work Guidance

  • Coordinate status endpoint behavior with frontend fallback decisions.

Verification

  • Run pytest tests/test_kokoro_tts.py tests/test_paradee_tts.py tests/test_speech_plugin_split.py tests/test_framework_dependencies.py.
  • Smoke-test both engines, voice-picker transitions, speed, status, and disabled browser fallback after changes.

Child DOX Index

No child DOX files.