Text-to-Speech

POST /v1/audio/speech

Turn text into natural speech with one OpenAI-compatible endpoint. Send JSON, get audio bytes back — switch providers by changing modelonly. Nothing is written to disk: audio is streamed straight back to you.

Example Request

curl https://synthorai.io/v1/audio/speech \
  -H "Authorization: Bearer $SYNTHORAI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "tts-1",
    "input": "Hello from Synthorai.",
    "voice": "alloy"
  }' \
  --output speech.mp3

Python (OpenAI SDK)

from openai import OpenAI

client = OpenAI(
    api_key="$SYNTHORAI_API_KEY",
    base_url="https://synthorai.io/v1",
)

with client.audio.speech.with_streaming_response.create(
    model="tts-1",
    voice="alloy",
    input="Hello from Synthorai.",
) as response:
    response.stream_to_file("speech.mp3")

Request Body

Parameter	Type	Description
`model`*	string	Model ID. One of `tts-1`, `tts-1-hd`, `seed-tts-2.0`.
`input`*	string	The text to synthesise. Maximum 4096 characters.
`voice`	string	Voice to speak in. Defaults to `alloy`. See Voices below.
`response_format`	string	Audio container: `mp3` (default), `opus`, `wav` or `pcm`.
`speed`	number	Speaking rate, 0.25–4.0. Defaults to 1.0.

Supported models

Which one should I use?

You need	Recommended models
Cheapest, good enough for most product voice	`google-tts-standard`
Chinese-first content, or you want many voice options	`qwen3-tts-instruct-flash` `seed-tts-2.0`
Everyday product voice, balancing quality against cost	`tts-1` `google-tts-neural2`
Best expressiveness — narration, companion devices, brand voice	`tts-1-hd` `google-tts-chirp3-hd`

Model	Price	Best for
`google-tts-standard`	$4 / 1M chars	High-volume notifications and prompts where cost dominates
`qwen3-tts-instruct-flash`	$11.50 / 1M chars	Chinese and multilingual content; upstream reports the billed character count
`tts-1`	$15 / 1M chars	General product voice, prompts, notifications
`google-tts-neural2`	$16 / 1M chars	Natural everyday voice across many languages
`tts-1-hd`	$30 / 1M chars	Narration, marketing, anything users listen to for a while
`seed-tts-2.0`	$30 / 1M chars	Chinese and multilingual content; 180+ voices
`google-tts-chirp3-hd`	$30 / 1M chars	Google's newest, most expressive voices

Three models sit at $30 — they differ in voice character and language coverage, not price. Try them on your own copy before committing.

ℹ

Switching providers costs you one line. Every model here speaks the same request and response shape. Providers differ wildly underneath (different auth, streaming formats, voice naming) — the gateway absorbs that so you only change model.

# Same request, different provider — only the model name changes.
-d '{"model": "tts-1",        "input": "...", "voice": "alloy"}'
-d '{"model": "seed-tts-2.0", "input": "...", "voice": "alloy"}'

Voices

OpenAI models accept the standard voices: alloy, echo, fable, onyx, nova, shimmer.

For seed-tts-2.0 the same six names work (they map to comparable voices), and you can also pass a native BytePlus voice ID directly for full control — for example zh_female_cancan_uranus_bigtts. The full list of 180+ native voices is in the BytePlus voice catalogue.

For qwen3-tts-instruct-flash the same six names work, or pass a native Qwen voice such as Cherry, Ethan, Nofish, Jada.

⚠

Google voices are tied to their price tier. Each Google model covers one tier, and the tier is part of the voice name. If you pass a native voice id such as en-US-Neural2-C you must use the matching model — google-tts-neural2 in that example. Cross-tier voices are rejected with a 502 rather than silently billed at the wrong rate.

Audio formats

Set response_format to pick the container. Default is mp3: mp3, opus, wav, pcm.

Not every provider supports every container; unsupported values fall back to MP3 rather than failing.

Billing

Charged per input character at the provider's official list price — no markup. One character = one Chinese character = one letter = one punctuation mark = one space. Failed requests are never charged.

The audio you get back is not metered separately; only the text you send in is.

Limits

Maximum 4096 characters per request. Longer input is rejected before anything is charged — split it client-side and concatenate the audio.
Speaking rate is controlled by speed (0.25–4.0, default 1.0). Values outside the provider's range are clamped, not rejected.

Errors

Response	Description
`400`	Input is empty, or longer than 4096 characters.
`401`	Missing or invalid API key.
`402`	Not enough quota for this request.
`502`	The model is not available on your account, or the upstream provider failed. Nothing is charged.
`503`	No channel currently serves this model.