Introduce an in-memory LRU cache for TTS audio with configurable
size and TTL via TTS_CACHE_MAX_SIZE_BYTES and TTS_CACHE_TTL_MS.
Return X-Cache headers (HIT/MISS) and set route runtime to nodejs.
Cache key includes provider, model, voice, speed, format, text,
and optional instructions.
Normalize non-Kokoro multi-voice input to the first token while
preserving full voice string in the cache key. Default Deepinfra
model to hexgrad/Kokoro-82M when none is provided.
Fix Deepinfra Kokoro behavior by enforcing single-voice selection:
- ui: only enable multi-select when provider supports >1 voices
- voice utils: Deepinfra max voices set to 1
- tests: gate provider selection and multi-voice tests by CI and
increase timeout for stability
- Introduce voice utils (model detection, voice parsing/weights, limits)
- Enable Kokoro multi-voice strings across OpenAI/Deepinfra/custom providers
- Normalize non-Kokoro voices to single token for SDK calls
- Expose full Kokoro voice list for custom-openai Kokoro models
- Update TTS API to return audio as ArrayBuffer and improve logging
- Add multi-select UI for Kokoro voices with provider-based clamping
- Preserve Kokoro voice strings in TTSContext and coalesce restarts
- Merge multi-sentence quoted dialogue in NLP sentence splitter
- Update tests for single and multi-voice selection flows
Add comprehensive multi-provider TTS support enabling users to choose between OpenAI, Deepinfra, and custom OpenAI-compatible endpoints. Implement provider-specific voice management with automatic voice restoration per provider-model combination, and migrate package manager to pnpm for improved dependency handling.
Key changes:
- Add TTS provider selection (OpenAI, Deepinfra, custom-openai) in settings UI
- Implement provider-specific model and voice lists with dynamic fetching
- Add voice persistence per provider-model combination in savedVoices
- Support Deepinfra models: Kokoro-82M, Orpheus-3B, Sesame-1B with their voice libraries
- Migrate to pnpm with frozen lockfile for reproducible builds
- Update Docker configuration to use pnpm and Deepinfra API defaults
- Add migration logic for existing users to infer provider from stored baseUrl
- Update test helpers and Playwright configuration for Deepinfra API
- Add example docker-compose.yml with Kokoro-FastAPI integration
BREAKING CHANGE: Voice selection is now provider-model specific. Previously saved voices will be migrated to the new savedVoices structure, but users may need to reselect voices if switching providers.
