> ## 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.

# Stream Speech

> Generate speech audio with chunked transfer for the lowest possible time-to-first-byte.

## Overview

Streams synthesized audio to the client as it is generated, using `Transfer-Encoding: chunked`. Functionally equivalent to [Speak](/api-v1/post/speak) with `stream: true`, but mounted as its own canonical endpoint for parity with our internal services and to avoid prefix conflicts with the non-streaming handler.

Use this endpoint when you care about time-to-first-byte (live preview, voice assistants, IVRs). Use [Speak](/api-v1/post/speak) when you want to buffer the full audio before delivering it.

<Note>
  Requires a Bland TTS voice (`BTTS`, `BTTS_V2`, or `BTTS_V3`). Other voice services are not supported on this endpoint.
</Note>

<Note>
  Every generation is automatically stored and retrievable via [List TTS Generations](/api-v1/get/speak-samples) and [Get TTS Generation](/api-v1/get/speak-samples-id), the same as the non-streaming endpoint.
</Note>

***

## Headers

<ParamField header="authorization" type="string" required>
  Your API key for authentication.
</ParamField>

***

## Body Parameters

<ParamField body="text" type="string" required>
  The text to synthesize. Maximum 5,000 characters per request. Supports pause markers in the form `<|N|>` (0.1-10.0 seconds).
</ParamField>

<ParamField body="voice_id" type="string" required>
  ID of the Bland TTS voice to use. Pass either the voice UUID from [List Voices](/api-v1/get/voices) or a curated voice name.
</ParamField>

<ParamField body="output_format" type="string" default="pcm_44100">
  Audio container/sample rate.

  * `pcm_8000`, `pcm_16000`, `pcm_24000`, `pcm_44100`, `ulaw_8000`
</ParamField>

<ParamField body="language" type="string">
  Language code. Defaults to the voice's primary language.
</ParamField>

<ParamField body="consistency" type="number">
  V1: float 0.0-1.0 (higher = more consistent). V2/V3: integer 1-32 (lower = more consistent).
</ParamField>

<ParamField body="expressiveness" type="number">
  V1 only. Float 0.0-1.0.
</ParamField>

<ParamField body="boost" type="integer">
  V2/V3 only. `0` or `1`.
</ParamField>

***

## Streaming response format

The response is a single WAV file delivered in chunks:

1. **First 44 bytes**, a standard WAV header with **placeholder sizes**: bytes 4-7 (RIFF chunk size) and 40-43 (data chunk size) are both filled with `0xFFFFFFFF` because the final length is not yet known.
2. **Subsequent chunks**, raw PCM16 audio data, written as it is synthesized.
3. **After the stream closes**, the client patches the WAV header in place: bytes 4-7 become the total file size minus 8, bytes 40-43 become the total data chunk size. Most decoders ignore the placeholder size and play the file fine without the patch, but tools that strictly validate the header will need it.

`Content-Type` is `audio/x-wav`. No `Content-Length` header is sent.

<ResponseField name="x-latency" type="string">
  Time in milliseconds from request to first audio byte.
</ResponseField>

<ResponseField name="x-cost" type="string">
  Cost in USD for the synthesis.
</ResponseField>

<ResponseField name="Transfer-Encoding" type="string">
  Always `chunked`.
</ResponseField>

<ResponseExample>
  ```http Response theme={null}
  HTTP/1.1 200 OK
  Content-Type: audio/x-wav
  Transfer-Encoding: chunked
  x-latency: 396
  x-cost: 0.001

  <chunked WAV with placeholder header sizes followed by PCM16 data>
  ```

  ```json Service Not Supported theme={null}
  {
    "data": null,
    "errors": [
      { "error": "Service Not Supported", "message": "This endpoint only supports Bland's Beige Voices" }
    ]
  }
  ```
</ResponseExample>

***

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