StepAudio 3 TTS | EmpirioLabs AI Docs

StepFun · Audio Generation

POST /v1/audio/speech

Human-level speech synthesis across six languages, with 12 system voices, natural-language delivery direction, and low-latency streaming.

At a glance

Field	Value
Model id	`stepaudio-3-tts`
Model release date	2026-09-09
Input modalities	Text
Output modalities	Audio
Context window	-
Weight precision	-
Region	International
Features	text_to_speech, contextual_tts, voice_control, multilingual
Native inference	No
New	Yes
Supported endpoints	`POST /v1/audio/speech`
Alternate model ids	`stepfun/stepaudio-3-tts`

Pricing

Charge	Spec	Rate
Synthesis	per 10,000 characters	$0.36

Example request

curl https://api.empiriolabs.ai/v1/audio/speech \
  -H 'Authorization: Bearer $EMPIRIOLABS_API_KEY' \
  -H 'Content-Type: application/json' \
  -d '{"model": "stepaudio-3-tts", "input": "Hello from EmpirioLabs."}'

Parameters

Parameter	Type	Required	Default	Description
`input`	string	yes	-	Text to synthesize. Maximum 1,000 characters. Content in parentheses is treated as delivery direction and is not spoken. · Max: 1000
`voice`	enum	no	`"lively-girl"`	Official StepFun voice. Custom cloned voice IDs are also accepted via the API. · Allowed: `lively-girl`, `vibrant-youth`, `soft-spoken-gentleman`, `magnetic-voiced-male`, `elegantgentle-female`, `livelybreezy-female`, `zixinnansheng`, `Alfie`, `Ethan`, `Matthew`, `qingshunvsheng`, `tianrunnvsheng`
`instruction`	string	no	-	Global emotion or style guidance for the whole passage, in natural language. Maximum 500 characters. · Max: 500
`speed`	number	no	`1`	Speech speed. · Range: 0.5 – 2
`volume`	number	no	`1`	Output volume. · Range: 0.1 – 2
`language`	enum	no	`"en"`	Target language. Japanese, Korean, French, and Spanish are in preview. · Allowed: `en`, `zh`, `ja`, `ko`, `fr`, `es`
`response_format`	enum	no	`"mp3"`	Output audio format. · Allowed: `mp3`, `wav`, `flac`, `opus`, `pcm`
`sample_rate`	enum	no	`24000`	Output sample rate in Hz. · Allowed: `8000`, `16000`, `22050`, `24000`
`markdown_filter`	boolean	no	false	Strip Markdown syntax from the input before it is spoken.
`pronunciation_map`	string	no	-	Optional pronunciation overrides, as an object with a tone array of “written/spoken” rules.

Notes

Maximum input is 1,000 characters. Content in parentheses is treated as delivery direction and is not spoken. Use instruction for global emotion or style guidance, up to 500 characters. Twelve system voices are available, and a custom cloned voice ID is also accepted. Recognized languages are Chinese, English, Japanese, Korean, French, and Spanish, with languages other than Chinese and English in preview. Output formats are mp3, wav, flac, opus, and pcm at 8000, 16000, 22050, or 24000 Hz. Wav is returned as a complete file rather than streamed. This model does not support voice_label.

Machine-readable schema: GET https://api.empiriolabs.ai/v1/models/stepaudio-3-tts.