Audio speech create

post

https://api.macpaw.com/v1/audio/speech

Generates audio from text using a unified OpenAI-compatible interface, routed directly to the provider selected by model.

Streaming only. The response is always a Server-Sent Events stream; stream_format may be omitted or set to sse, and audio (a single buffered audio file) is rejected with 400.

Required fields (this endpoint):

Field	Description
`model`	`provider/model` (e.g. `openai/gpt-4o-mini-tts`, `elevenlabs/eleven_multilingual_v2`).
`input`	Text to synthesize.
`voice`	Provider voice: an OpenAI voice name (e.g. `alloy`) or an ElevenLabs `voice_id`.

Provider endpoints

Provider	Upstream
OpenAI	Create speech
ElevenLabs	Stream speech with timing
Response shape. By default every provider's stream is normalized to OpenAI `speech.audio.delta` / `speech.audio.done` events, so a client written against OpenAI works unchanged. Set `raw_provider_response` to `true` to receive the provider's own chunks instead — for ElevenLabs that is its native JSON stream, including the alignment data the OpenAI events cannot carry.

Recent Requests

Time	Status	User Agent
Retrieving recent requests…

Loading…

Body Params

Request body requires model, input, and voice. Every other field is provider-dependent: which parameters a model accepts, their bounds, and their enums are defined per provider/model in Platform AI Model and validated before the provider is called. Consult the upstream documentation for their meaning:

model

string

required

Model in provider/model format. Only openai/* and elevenlabs/* models generate speech.

input

string

required

Text to synthesize. The maximum length is the resolved model's own character cap (e.g. 4096 for openai/gpt-4o-mini-tts, 40000 for elevenlabs/eleven_flash_v2_5).

voice

string

required

Voice to speak with: an OpenAI voice name (e.g. alloy) or custom voice object, or an ElevenLabs voice_id.

response_format

string

Audio format, in the resolved provider's own values — OpenAI codec names (mp3, opus, aac, flac, wav, pcm) or ElevenLabs codec_samplerate_bitrate tokens (e.g. mp3_44100_128, pcm_16000). Omitted uses the provider's default.

stream_format

string

enum

Only sse is supported: speech is always streamed. audio (a single buffered file) is rejected with 400.

Allowed:

raw_provider_response

boolean

Defaults to false

When true, relays the provider's own response chunks unmodified instead of normalizing them to speech.audio.* events. It is a platform control, never forwarded to the provider and never validated as a model parameter. OpenAI models already stream the OpenAI shape, so it only changes the response for other providers.

Headers

string

enum

Defaults to application/json

Generated from available response content types

Allowed:

Responses

200Stream of audio events. By default: SSE events named speech.audio.delta (base64 audio chunk) followed by a single speech.audio.done. With raw_provider_response: true the provider's own stream is relayed unmodified, so its Content-Type and chunk shape apply instead.

401Invalid token

402Insufficient credits

403Forbidden — the caller may not use the requested model: it is not registered with active parameter rules, or their team is not entitled to it. Both are the same refusal from the caller's side.