Text to speech

POST /v1/audio/speech is OpenAI-compatible speech synthesis. The response is binary audio with a content type derived from response_format.

curl "$SONIX_URL/v1/audio/speech" \
  -H "Authorization: Bearer $SONIX_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "kokoro-82m",
    "input": "Hello from sonix.",
    "voice": "af_heart",
    "response_format": "wav",
    "speed": 1.0
  }' --output hello.wav

Request fields

Field	Required	Notes
`model`	yes	A speech-capable model ID. See Models.
`input`	yes	The text to synthesize.
`voice`	yes	A voice ID or OpenAI-style name (`alloy`, …), or an object like `{ "id": "af_heart" }`. See Voices.
`response_format`	no	`mp3` (default), `opus`, `aac`, `flac`, `wav`, or `pcm`.
`speed`	no	`0.25` to `4.0`. Defaults to `1.0`.
`instructions`	no	Forwarded to the model as synthesis metadata. Not delivered to `qwen3-tts-0.6b`; see below.
`stream_format`	no	`audio` (default). `sse` is rejected until event streaming ships.

wav responses are wrapped in a WAV container server-side; other formats are returned as produced by the model. Unsupported parameters are rejected with a validation error rather than silently ignored.

Format support varies by model. qwen3-tts-0.6b produces pcm and wav only: an omitted response_format returns wav there, an explicit mp3, opus, aac or flac returns a validation error. On that model speed is accepted but currently unapplied; the x-zzz-speed-applied response header reports the speed actually used. instructions is not delivered to that model either; only input, voice, response_format and speed reach it.

Using the OpenAI SDK

Any OpenAI SDK works by overriding the base URL:

from openai import OpenAI

client = OpenAI(base_url=f"{SONIX_URL}/v1", api_key=SONIX_API_KEY)

speech = client.audio.speech.create(
    model="kokoro-82m",
    voice="af_heart",
    input="Hello from sonix.",
    response_format="wav",
)
speech.write_to_file("hello.wav")