> ## 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 Deepgram

> Move Deepgram's streaming Speak protocol to Bland realtime turns.

Deepgram and Bland both return raw binary audio over a persistent WebSocket. Most of the migration is a control-message rename plus one important change: Bland handles text buffering without requiring `Flush` after each useful fragment.

Start with the [Realtime TTS Quickstart](/tts/realtime-quickstart).

## Message mapping

| Deepgram streaming TTS                                 | Bland `/v2/tts/ws`                                                |
| ------------------------------------------------------ | ----------------------------------------------------------------- |
| Voice model and audio settings in URL query parameters | Voice UUID and audio settings in `init`                           |
| `{ "type": "Speak", "text": ... }`                     | `{ "type": "speak", "context_id": ..., "text": ... }`             |
| `{ "type": "Flush" }`                                  | Usually unnecessary; use `end_of_turn` only when the LLM finishes |
| `{ "type": "Clear" }`                                  | `cancel`, or start a replacement context to preempt               |
| `{ "type": "Close" }`                                  | `close` after the final `utterance_end`                           |
| `Flushed`                                              | No direct event; use `utterance_end` for the turn terminal        |
| Binary audio                                           | Binary audio                                                      |

```js theme={null}
for await (const token of llmTextStream) {
  ws.send(
    JSON.stringify({
      type: "speak",
      context_id: turnId,
      text: token,
    }),
  );
}

ws.send(JSON.stringify({ type: "end_of_turn", context_id: turnId }));
```

Do not translate every Deepgram `Flush` into a new Bland turn. Keep one Bland context for the entire LLM response. Bland begins synthesis at natural boundaries while more tokens continue to arrive.

Deepgram's `Clear` acknowledges cleared server state. Bland reports `utterance_end` with reason `cancelled` or `preempted`. In both cases, clear any unplayed audio in your local playback queue.

See the [Deepgram streaming TTS reference](https://developers.deepgram.com/reference/text-to-speech/speak-streaming) and [Bland realtime concepts](/tts/realtime-concepts).

***

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