GPT-Live Real-Time Voice Conversations

GPT-Live supports listening and speaking simultaneously, bidirectional subtitles, and background tasks. ChatGPT voice calls in Studio use gpt-live-1 to process speech, while complex questions continue to be handled by the currently selected chat model.

¶ Connection

Use a Platform API Token to connect to wss://realtime.acedata.cloud/v1/live/sessions. Use wss://realtime.acedata.cloud/aichat2/live for Studio or chat service credentials. The server authenticates through Authorization: Bearer <token>; browsers pass credentials through the WebSocket subprotocol to avoid writing them in the URL:

const socket = new WebSocket('wss://realtime.acedata.cloud/aichat2/live', [
  'live', `acedata-token.${token}`
]);
socket.onopen = () => socket.send(JSON.stringify({
  type: 'session.start',
  session: {
    model: 'gpt-live-1',
    instructions: 'Answer concisely and friendly, using the user’s language. Hand complex questions to the background.',
    audio: { format: { type: 'audio/pcm', rate: 24000 }, output: { voice: 'marin' } },
    delegation: { type: 'client' },
    store: false
  }
}));

After waiting for session.started, continuously send session.input_audio.append, where audio is the Base64 of raw 24 kHz, mono, 16-bit little-endian PCM audio, without a WAV file header. Send audio at the actual sampling rate, and keep silent segments continuous as well.

Play session.output_audio.delta.delta in order. User and assistant subtitles come from session.input_transcript.delta and session.output_transcript.delta, respectively; retain the delta, start_ms, and end_ms of each segment; subtitles from both sides may grow simultaneously. Live does not have an end event for each individual voice response; the “responding” state should be updated based on the local playback queue.

¶ Background Tasks

The client delegation mode is currently supported. After receiving session.delegation.created, the application calls its own backend based on voice subtitles and saved context, and returns the result through session.thinking.append or session.commentary.append, carrying the corresponding delegation_id and plain-text content of up to 500 tokens. Results should come from completed backend operations; interrupted speech does not mean that a background task has been canceled.

Managed Responses delegation, recording storage, session.update, session recovery, or automatic reconnection are not supported for now. The model is fixed to gpt-live-1, audio is fixed to 24 kHz PCM16, and the voice is selected when the call starts. The old Realtime API can still be used.

¶ Billing and Ending Calls

Voice is billed by active session duration: 0.01 Credits/second, or 0.6 Credits/minute, with no rounding up to full minutes. Time when the user is speaking, the assistant is speaking, there is silence, both parties are silent, or background tasks are pending is all billable. Background model or tool calls are billed separately according to their respective APIs. The USD value of Credits depends on the top-up plan; refer to the current console pricing.

session.usage.updated.usage.seconds is cumulative usage. For example, if 12 seconds are received first and then 15 seconds, the total is 15 seconds. The platform only charges for additional duration and does not double-count snapshots.

When ending a call, send session.close and continue receiving until session.closed is received. Then close the connection and release the microphone. A forced disconnection may prevent final usage from being confirmed, which should be clearly reported in the interface. The connection ends if the balance is insufficient, billing fails, or the session limit is reached. A single call is limited to 30 minutes by default.

Parameter errors are returned through the error event; error.client_event_id corresponds to a failed command. If startup fails, stop recording and display the error to avoid repeated requests incurring additional charges.