Summary

• add POST /v1/audio/speech with the standard model, input, voice, response_format, and speed request fields
• route common OpenAI TTS model aliases through Kokoro while preserving native Speech Swift model and voice aliases
• return WAV by default or headerless 24 kHz mono PCM16 when requested
• return OpenAI-shaped error envelopes and reject unsupported compressed formats explicitly
• document the HTTP audio contract and expose the route in server startup output

Architecture fit

The route is a thin protocol adapter over the existing model registry and dispatchSynthesize path. It adds no model runtime or codec dependency. Voice and speed parameters are carried through the shared dispatcher only where the selected engine supports them.

Compatibility notes

This endpoint supports WAV and PCM output. Explicit MP3, Opus, AAC, and FLAC requests return a clear 400 response rather than mislabeled audio. Generic OpenAI voice names select the local engine default; native Kokoro and Qwen3-TTS voice names pass through.

Tests

• swift test --skip E2E
• swift test --filter OpenAISpeechTests
• swift test --filter E2EOpenAISpeechTests