Ultravox is a platform for real-time voice agents. Kova plugs in through Ultravox’s generic external-TTS adapter: you point an agent straight at the Kova API, no Kova-side setup required, so this works today.
Set an externalVoice of type generic on call (or agent) creation, pointed at the Kova /v1/tts endpoint. Kova returns the audio base64-encoded inside a JSON body, so jsonAudioFieldPath and jsonByteEncoding tell Ultravox how to read it.
body.text: keep it as "{text}". Ultravox replaces it with the agent’s text on each turn.body.voice: any speaker id from /v1/tts/speakers, for example ash.body.response_format: encoding: "pcm" returns raw 16-bit mono little-endian PCM. Keep sample_rate equal to responseSampleRate.responseMimeType / jsonAudioFieldPath: Kova returns { "audio": "<base64>" } with Content-Type: application/json, so Ultravox reads the audio field and base64-decodes it. The application/jsonl value is Ultravox’s own parser selector — it does not need to match Kova’s actual Content-Type. This exact config is validated end to end against Ultravox Realtime.headers.x-api-key: your Kova API key from platform.kova.ai. Ultravox sends it on every request; Kova validates and bills it.
externalVoice is mutually exclusive with the voice field, set one or the other. Each config maps to one Kova voice; create one externalVoice per voice you want to offer.
Billing and errors
Usage is billed to the API key in the x-api-key header, the same as direct API calls. An invalid key returns 401; an account out of credits returns 402. Top up at platform.kova.ai/dashboard/billing. 
