StepAudio 3 Realtime | EmpirioLabs AI Docs

StepFun · Audio Generation

WEBSOCKET /v1/realtime

Speech in, speech out, over one socket. Interruptible mid-reply, with voice activity detection deciding when to answer.

At a glance

Field	Value
Model id	`stepaudio-3-realtime`
Model release date	2026-08-06
Input modalities	Audio, Text
Output modalities	Audio, Text
Context window	256K
Weight precision	-
Max output tokens	131,072
Region	International
Features	realtime, speech_to_speech, audio_in, audio_out, voice_control
Native inference	No
New	Yes
Supported endpoints	`WEBSOCKET /v1/realtime`
Alternate model ids	`stepaudio-3-realtime-preview`, `stepfun/stepaudio-3-realtime`

Pricing

Charge	Spec	Rate
Input	per 1M prompt tokens	$1.50
Output	per 1M generated tokens	$10.00
Implicit cache read	per 1M cached input tokens	$0.30

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-realtime", "input": "Hello from EmpirioLabs."}'

Parameters

Parameter	Type	Required	Default	Description
`voice`	enum	no	-	Session voice, set with session.update before the model produces audio. It cannot be changed once the model has spoken, and any other value is rejected. · Allowed: `soft-spoken-gentleman`, `magnetic-voiced-male`, `vibrant-youth`, `lively-girl`, `livelybreezy-female`, `elegantgentle-female`, `zixinnansheng`
`instructions`	string	no	-	System instructions for the session: persona, speaking style, and boundaries.
`modalities`	string	no	`["text", "audio"]`	Response modalities for the session.
`input_audio_format`	enum	no	`"pcm16"`	Encoding of the audio you send. 16-bit PCM. · Allowed: `pcm16`
`output_audio_format`	enum	no	`"pcm16"`	Encoding of the audio the model returns. 16-bit PCM. · Allowed: `pcm16`
`turn_detection`	string	no	-	Server-side voice activity detection settings, which decide when the model starts and stops replying.

Notes

Full-duplex voice over a WebSocket at wss://api.empiriolabs.ai/v1/realtime?model=stepaudio-3-realtime, authenticated with the ordinary Authorization Bearer header. Audio is pcm16 in and out. The model listens while it speaks, so it can be interrupted mid-reply, and server-side voice activity detection decides when to answer. Set voice with session.update before the first audio; it cannot be changed once the model has spoken, and only the seven listed voices are accepted. Conversation is Chinese and English only. Each completed turn is billed on its own from the usage the model reports. This model is in preview and its capabilities may change.

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