[remsky]

OpenAI Realtime compatible TTS on /v1/realtime, for Realtime TTS clients (wyoming_openai etc). TTS only: response.create text input, 24 kHz PCM output. OpenAI's Speech models shut down Jan 2027 and point users to Realtime.

• Voice names map the same as on the Speech endpoint, as a string or {"id": ...}
• voice_aliases, lang_code, volume_multiplier, normalization_options work on all three websocket routes (shared SpeechOptions), no voice tags
• response.cancel stops a response. Audio input and conversation events get an error back
• Same ENABLE_WEBSOCKET_TTS switch and API_KEYS auth as the ElevenLabs/Cartesia routes
• Client tests use the real openai SDK (3.44.0, test-only dependency)

ElevenLabs route now checks the voice on the first context message instead of at connect, since aliases arrive with it. Bad voices still close with 1008.

Testing

• uv run pytest passes locally (API changes)
• npm run test:web passes locally (web player changes)
• New or changed behavior has a test
• Ran in Docker (npm run cpu:up / gpu:up), for inference or docker changes
• CHANGELOG.md entry, if users will notice the change