> ## Documentation Index
> Fetch the complete documentation index at: https://docs.bland.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Migrate from Cartesia

> Map Cartesia WebSocket contexts and continuations to Bland turns and preemption.

Both APIs use a persistent WebSocket and a context ID per conversational turn. Bland makes barge-in the default: a new context replaces the active context instead of running concurrently.

Start with the [Realtime TTS Quickstart](/tts/realtime-quickstart), then translate the protocol as follows.

## Message mapping

| Cartesia WebSocket                                    | Bland `/v2/tts/ws`                              |
| ----------------------------------------------------- | ----------------------------------------------- |
| Generation request with `context_id` and `transcript` | `speak` with `context_id` and `text`            |
| `continue: true`                                      | Keep sending `speak` with the same `context_id` |
| `continue: false` or final empty transcript           | `end_of_turn`                                   |
| `{ context_id, cancel: true }`                        | `{ type: "cancel", context_id }`                |
| Base64 `chunk.data` in JSON                           | Raw binary audio frame                          |
| `done: true` for a context                            | `utterance_end`                                 |
| Repeated model, voice, language, and output fields    | One `init` message for the connection           |

```js theme={null}
ws.send(
  JSON.stringify({
    type: "speak",
    context_id: "turn-9",
    text: "Thanks for calling, ",
  }),
);
ws.send(
  JSON.stringify({
    type: "speak",
    context_id: "turn-9",
    text: "how can I help today?",
  }),
);
ws.send(JSON.stringify({ type: "end_of_turn", context_id: "turn-9" }));
```

## Important differences

* Cartesia can multiplex multiple contexts over one socket. Bland has exactly one active turn. Sending a different `context_id` preempts the old one.
* Cartesia echoes the context ID on audio JSON. Bland brackets raw binary frames with `utterance_start` and `utterance_end`, so track the active context locally.
* Bland handles token buffering. You do not need to port `max_buffer_delay_ms` or manual flush IDs.
* Bland does not currently emit word or phoneme timestamps.
* Bland cancellation stops forwarding the active turn. Clear locally queued audio as well because delivered frames cannot be recalled.

See the [Cartesia WebSocket reference](https://docs.cartesia.ai/api-reference/tts/websocket), [Cartesia context guide](https://docs.cartesia.ai/use-the-api/tts-websocket/contexts), and [Bland turn concepts](/tts/realtime-concepts#the-connection-and-turn-model).

***

Docs for agents: [llms.txt](/llms.txt)