- Add PluginManager singleton for plugin discovery and engine caching
- Add CompatBackend adapter wrapping Engine/EngineSession into old create_backend() API
- Update tts_plugin/__init__.py with public exports
- Migrate preview.py and its test to use compat.create_backend
- Add integration and plugin manager contract tests
Implement plugin loading and validation:
- loader.py: discover, import, validate plugins
- validate PLUGIN_MANIFEST, MODEL_REQUIREMENTS, create_engine
- validate api_version compatibility (major must match)
- validate capabilities (reject unknown)
- diagnostic messages for all error cases
- no partial registration after error
Test plugins:
- fake_plugin: minimal valid plugin for testing
- missing_manifest: no PLUGIN_MANIFEST
- invalid_api_version: major version mismatch
- invalid_capabilities: unknown capabilities
- missing_create_engine: no create_engine function
- import_error: raises ImportError during import
- missing_model_requirements: no MODEL_REQUIREMENTS
39 new tests covering all loader functionality.
Create public API structure for TTS Plugin Architecture:
- types.py: immutable value objects (AudioFormat, Duration, VoiceSelection, etc.)
- errors.py: EngineError hierarchy (7 typed exceptions)
- manifest.py: plugin manifest dataclasses (PluginManifest, EngineManifest, etc.)
- engine.py: Engine and EngineSession protocols
- capabilities.py: optional capability interfaces (VoiceLister, PreviewGenerator, etc.)
- host_context.py: HostContext and HttpClient protocol
- plugin.py: plugin contract (create_engine signature)
- __init__.py: public API exports
All interfaces are fully defined but contain no business logic.
API is frozen and ready for implementation in subsequent PRs.
- Add resolve_backend_for_voice() to TTSBackendRegistry
- Add module-level wrapper resolve_backend_for_voice()
- Simplify _infer_provider_from_spec() to use registry API
- Simplify _supertonic_voice_from_spec() to only normalize
- Add 11 test cases for the new method
Resolution rules:
1. Empty spec -> fallback
2. Kokoro formula (* or +) -> kokoro
3. Exact voice ID match -> backend id
4. Unknown voice -> fallback
Add TTSBackendRegistry.is_registered() and module-level
is_registered_backend() to validate backend IDs dynamically.
Replace all Category A hardcoded sets (validation-only) in
voice_profiles, api routes, conversion_runner, and form utils.
Make the Kokoro voice list an internal implementation detail of the
backend instead of a shared constant. The rest of the project already
accesses voices via get_metadata('kokoro').voices.
- Move VOICES_INTERNAL from constants.py to kokoro.py as _VOICES_INTERNAL
- Update tests to use get_metadata('kokoro').voices instead of importing
the constant directly
Replace direct VOICES_INTERNAL imports with get_metadata('kokoro').voices
from tts_backend_registry in all PyQt modules:
- abogen/pyqt/gui.py
- abogen/pyqt/predownload_gui.py
- abogen/pyqt/voice_formula_gui.py
- Add get_default_voice() helper to tts_backend_registry
- Replace all VOICES_INTERNAL imports in WebUI with get_metadata().voices
- Replace all DEFAULT_SUPERTONIC_VOICES imports in conversion_runner with get_metadata().voices
- Remove unused VOICES_INTERNAL import from voices.py
Core modules (voice_profiles, voice_formulas, voice_cache) already used
get_metadata(). This completes the WebUI migration.
Replace direct imports of VOICES_INTERNAL and DEFAULT_SUPERTONIC_VOICES
in voice_profiles, voice_formulas, and voice_cache with get_metadata()
from TTSBackendRegistry. Adds get_metadata() top-level function to
tts_backend_registry as symmetric counterpart to register_backend() and
create_backend().
Add voices field to TTSBackendMetadata so each backend's supported
voice list is part of its metadata rather than external constants.
- Add voices: tuple[str, ...] = () to TTSBackendMetadata
- Create _KOKORO_METADATA / _SUPERTONIC_METADATA as single source
of truth for both metadata property and registry registration
- Update KokoroBackend.get_available_voices() to use self.metadata.voices
- Update SupertonicBackend.get_available_voices() to use self.metadata.voices
- Add tests for voices field, metadata voice content, and unified instance identity
- Create KokoroBackend class implementing TTSBackend protocol
- Move all KPipeline interaction inside KokoroBackend
- Update LoadPipelineThread to create backend via create_backend()
- Update ConversionThread and VoicePreviewThread to accept backend
- Replace np_module/kpipeline_class parameters with single backend
- Add 24 unit tests for KokoroBackend
- KPipeline is now an internal implementation detail of KokoroBackend
Encapsulate SupertonicPipeline as an internal detail of
SupertonicBackend. The factory create_supertonic_backend() now
returns a SupertonicBackend instance instead of a raw
SupertonicPipeline, satisfying the TTSBackend protocol with
metadata, synthesize, get_available_voices, get_supported_formats,
and get_info methods. Backward-compatible __call__ delegates to
the internal pipeline.
Migrate all preview/debug/conversion pipeline creation to use
TTSBackendRegistry.create_backend() instead of direct imports:
- debug_tts_runner._load_pipeline(): Kokoro via registry
- preview.get_preview_pipeline(): Kokoro via registry
- preview.generate_preview_audio(): Supertonic via registry
- voice.get_preview_pipeline(): Kokoro via registry
- conversion_runner._load_pipeline(): both backends via registry
- conversion_runner inline pipeline creation: both via registry
- test: update mock to target tts_backend_registry.create_backend
- Add create_kokoro_backend() factory in kokoro.py
- Add create_supertonic_backend() factory in supertonic.py
- Auto-discover backend modules in __init__.py via pkgutil
- Both backends register themselves on import
- Tests verify registration and factory callables
- TTSBackendRegistry class with register(), list_backends(), get_metadata(), create_backend()
- Global registry singleton with register_backend() and create_backend() convenience functions
- Unit tests for registry operations
Moved SupertonicPipeline from abogen/tts_supertonic.py to
abogen/tts_backends/supertonic.py and load_numpy_kpipeline from
abogen/utils.py to abogen/tts_backends/kokoro.py.
Git correctly detects the Supertonic file as a rename (R),
preserving full commit history.
- New package: abogen/tts_backends/
- __init__.py (package marker)
- supertonic.py (SupertonicPipeline, moved from tts_supertonic.py)
- kokoro.py (load_numpy_kpipeline, moved from utils.py)
- abogen/utils.py: re-exports load_numpy_kpipeline for backward compat
- All imports updated to new canonical paths
- Create TTSBackend abstract base class with minimal contract
- Implement KokoroTTSBackend that maintains existing behavior
- Update conversion_runner.py to use new interface
- No behavioral changes, GUI unchanged, no new features
fix: catch OverflowError in emit_text for very large numbers
Fixes#145
When text contains a very large number (e.g. a long decimal from a binary
hash), misaki's pipeline calls num2words() which raises OverflowError and
crashes the entire job. Wrap the segment iterator in try/except so the
chunk is skipped gracefully with a warning log instead of terminating.
The upstream project moved PyQt code to abogen/pyqt/ subdirectory, making the
original feature commits non-mergeable. This commit re-applies both features
to the new file locations.
Voice Marker feature (<<VOICE:voice_name>> syntax):
- subtitle_utils.py: Added _VOICE_MARKER_PATTERN, _VOICE_MARKER_SEARCH_PATTERN,
validate_voice_name(), split_text_by_voice_markers() (with valid/invalid counts)
- pyqt/conversion.py: Added load_voice_cached(), voice marker pre-processing before
chapter loop, inner voice segment loop wrapping spaCy+TTS block, updated imports
- pyqt/gui.py: Added Insert Voice Marker button and insert_voice_marker() to TextboxDialog
Word Substitution feature (text preprocessing before TTS):
- word_substitution.py: New module (word replacements, ALL CAPS, numerals, punctuation)
- pyqt/conversion.py: apply_word_substitutions() call after clean_text()
- pyqt/gui.py: WordSubstitutionsDialog, word_sub_combo, Settings button,
on_word_sub_changed(), show_word_sub_dialog(), config persistence, queue restore
- pyqt/queued_item.py: 6 new word substitution fields
- pyqt/queue_manager_gui.py: 6 fields added to OVERRIDE_FIELDS and get_current_attributes()
Note: num2words>=0.5.13 was already added to pyproject.toml by upstream.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
The app was crashing after importing a .txt and clicking convert because of a missing import. Fixed the imports and removed the legacy abogen.conversion module which doesn't seem necessary anymore.
The 'abogen' command now launches the PyQt6 desktop GUI which requires
display libraries not available in Docker containers. Changed the Docker
CMD to use 'abogen-web' which launches the Flask web UI.
Artem Akymenko