PocketTTS 2.0

[Shackless]

fixes #377 #389

[Copilot]

Pull request overview

Upgrades the local TTS stack to PocketTTS 2.0 (multilingual + model selection/quantization) and adds supporting runtime APIs (model status, voice preloading/precompute, voice preprocessing), while also improving STT model download/init UX and introducing a global spoken_language setting that feeds into prompt formatting and STT/TTS defaults.

Changes:

• Migrate PocketTTS integration to v2.0 with model IDs, quantization (torchao), LRU voice-state caching, and background (pre)compute/preload flows.
• Add new core API routes for PocketTTS status/models, voice preloading/precompute, and an upload-based voice preprocessing pipeline.
• Add spoken_language setting + migration and inject {language_instruction} into the default system prompt; set PocketTTS as the default TTS provider.

Reviewed changes

Copilot reviewed 15 out of 16 changed files in this pull request and generated 4 comments.

Show a summary per file

File	Description
WingmanAiCore.spec	Includes torchao artifacts in PyInstaller build to support PocketTTS int8 quantization.
wingman_core.py	Adds PocketTTS status/models/preload/precompute/preprocess endpoints; defers PocketTTS init; warms voice cache on tower init.
templates/configs/settings.yaml	Updates PocketTTS defaults (model + quantize) and adds spoken_language.
templates/configs/defaults.yaml	Injects {language_instruction} and switches default tts_provider to pocket_tts.
services/voice_preprocessing.py	New audio preprocessing utility used for voice cloning uploads.
services/stt_provider_manager.py	Downloads FasterWhisper models via ModelDownloader with UI status callbacks; status propagation on provider switches.
services/settings_service.py	Adds spoken_language cascade logic and status callbacks for runtime STT provider switching.
services/migrations/migration_311_to_312.py	Migrates PocketTTS settings to new model/quantize fields, adds spoken_language, resets system prompt, switches default TTS provider.
services/file.py	Adds get_pocket_tts_models_dir() helper.
services/context_builder.py	Provides {language_instruction} substitution based on spoken_language.
services/audio_player.py	Best-effort saving of generated preview audio to disk for later use.
requirements.txt	Bumps pocket-tts to 2.0.0 and adds torchao.
providers/pocket_tts.py	Major refactor for PocketTTS v2: model IDs, custom model discovery, per-model cache tagging, precompute/preload/status, and concurrency locks.
providers/parakeet.py	Adds language cascade behavior (wingman override → global setting).
api/interface.py	Updates PocketTTS settings schema and adds spoken_language + preload response model.

Comments suppressed due to low confidence (1)

providers/pocket_tts.py:286

• unload_model() mutates self.model and clears voice_cache without acquiring _model_swap_lock. Because synthesis runs in a worker thread under _model_swap_lock, a concurrent call to unload_model() (e.g., via the stop endpoint or a settings change) can delete/NULL the model mid-generation and crash. Make load_model()/unload_model() take _model_swap_lock internally (or provide a single public ‘swap’ method that always holds the lock) so all entrypoints are safe.

    def unload_model(self):
        """Unload the model to free resources."""
        if self.model:
            del self.model
            self.model = None

        # Explicitly clear CUDA cache if using GPU to free GPU memory
        if torch.cuda.is_available():
            torch.cuda.empty_cache()

        self.voice_cache.clear()

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.