UpVoot — Voice AI Agents for Indian Businesses

The UpVoot voice protocol is designed to be minimal. Most integrations only need to stream binary audio frames — the JSON control messages are optional signals you can listen to for logging, UI updates, or custom logic.

Message types

Binary frames

Every binary WebSocket frame is a chunk of raw PCM audio. There is no framing header — the entire payload is audio data.

Client → server: Caller audio (microphone input)
Server → client: Agent audio (to be played to the caller)

You can send chunks of any size. Smaller chunks (e.g. 20ms = 640 bytes) reduce latency; larger chunks (e.g. 100ms) reduce overhead. We recommend 20–40ms chunks.

Call state machine

           WebSocket open
                │
                ▼
      ┌─────────────────┐
      │   GREETING      │  Agent speaks greeting message
      └────────┬────────┘
               │  audio_end
               ▼
      ┌─────────────────┐
      │   LISTENING     │  Waiting for caller to speak
      └────────┬────────┘
               │  caller speaks (binary audio)
               ▼
      ┌─────────────────┐
      │   PROCESSING    │  turn_start — LLM generating response
      └────────┬────────┘
               │  audio_start
               ▼
      ┌─────────────────┐
      │   SPEAKING      │  Agent speaking (binary audio to client)
      └────────┬────────┘
               │  audio_end  (or barge-in)
               ▼
      ┌─────────────────┐
      │   LISTENING     │  Loop until max turns or end_call
      └─────────────────┘

Barge-in handling

If the caller speaks while the agent is in SPEAKING state, the platform automatically:

Stops the current agent audio stream
Sends audio_end
Begins processing the caller's interruption

Your telephony provider does not need to do anything special — just keep streaming caller audio even while receiving agent audio.

Call termination

The call can end in several ways:

Trigger	Who	Behavior
Client closes WebSocket	Your provider	Call ends immediately; call is logged
`end_call` message	Client	Agent says a goodbye, then closes
Max turns reached	Server	Server sends `call_end` with `reason: "max_turns"`
Balance exhausted	Server	Server sends `call_end` with `reason: "balance_exhausted"`
Silence timeout	Server	Server sends `call_end` with `reason: "silence_timeout"`

Minimal listener (JavaScript)

ws.addEventListener("message", (event) => {
  if (event.data instanceof ArrayBuffer) {
    // Binary: agent audio — send to caller
    playOrForward(event.data);
    return;
  }

  const msg = JSON.parse(event.data);
  switch (msg.type) {
    case "transcript":
      console.log(`[${msg.role}]: ${msg.text}`);
      break;
    case "call_end":
      console.log("Call ended:", msg.reason);
      ws.close();
      break;
    case "error":
      console.error("Platform error:", msg.code, msg.message);
      break;
  }
});