Start a speech-to-speech session (audio only, no video)

Reserves a speech session with an avatar: a realtime speech-to-speech conversation over a WebSocket, without video. The answer carries the WebSocket url and a short-lived token: connect to url within 60 seconds (tokenExpiresAt), presenting the token either as Authorization: Bearer <token> or as the WebSocket subprotocols atmee-live and atmee-token.<token>. The token only admits the connection. The session ends at its endsAt (the reservation plus actualDurationSeconds), whenever the connection opened. The first connection holds the session for good: no other connection can use it, so a client that lost its connection starts a new session.

avatarId is required: the conversation uses that avatar's persona (name, personality, facts, self-introduction) and always its voice; an avatar without a voice answers 422 avatar_has_no_voice.

settings pins session.start fields from your backend: the speech server applies them over the client's, refuses a client session.start / session.update that sets a pinned field (naming it), and leaves them out of the session it echoes. Unpinned fields stay the client's to set. Pinned persona fields override the avatar's. Appends (session.instructions.append and the like) stay allowed: they add to the pinned instructions, never replace them.

Billing runs by the minute from the speech server's first usage report (within 30 s of the connection) until the session ends, at the plan's speech rate. A session nobody connects to costs nothing and is closed (failed, no_user_joined) by the first read, or start of the account, 30 seconds after its token expired, otherwise by the stale-session cleanup: 120 seconds without a heartbeat (the reservation sets the first), checked every minute.

Who pays. Always the account that owns the avatar:

• API key (X-Api-Key) — your backend, billed to the key's account. Requires the sessions scope; an avatar-scoped key may only start sessions on its avatar (it must name it).
• The owner's login (Authorization: Bearer <jwt>, where the deployment accepts user tokens) — billed to the account the user acts for (X-Account-Id when there are several). An avatar of another account answers 403 avatar_not_allowed.
• No credential — 401. There is no anonymous or visitor-paid speech session.

With both an API key and a bearer token, the key's start applies, as on POST /v1/sessions. Read and end a session with either credential of its account.