What is StepAudio 3?
StepAudio 3 is five models under one name: TTS for expressive speech, ASR Max for transcription, Realtime for full-duplex voice agents that can call tools, Gen for speech mixed with sound effects and music, and Music for full songs. All of them run on StepFun's API and in its Voice Studio web app, and the line sits above StepAudio 2.5, which StepFun still sells.
TTS and ASR Max are billed per use now; Realtime, Gen and Music are free previews that StepFun will replace with paid versions. Language support is narrow: six languages for speech and transcription, four of them in preview, and only Chinese and English for Realtime.
What changed from StepAudio 2.5?
Better transcription, cheaper speech, a voice agent that thinks while it talks, and two new models for sound and music; only ASR Max costs more.
- Speech price
- StepAudio 3 TTS costs $0.36 per 10,000 characters, against $0.85 for StepAudio 2.5 TTS.
- Transcription
- StepFun reports a 0.57% error rate on its context test, down from 1.59% for StepAudio 2.5 ASR, but ASR Max costs $0.40 an hour against $0.022.
- Voice agents
- Realtime adds thinking while it speaks and runs tools in the background while the conversation goes on.
- New models
- Gen and Music have no StepAudio 2.5 counterpart.
Who is StepAudio 3 best for?
StepAudio 3 is best for developers who want speech synthesis, transcription, real-time voice agents and music generation from one API, especially for Chinese and English audio.
Real-time voice agents in Chinese or English
StepAudio 3 Realtime holds full-duplex conversations, copes with interruptions, reasons while it speaks and can call tools.
Transcripts full of jargon
ASR Max uses context and domain knowledge to catch names, drug names and technical terms across more than 20 fields.
Expressive narration
StepAudio 3 TTS adds laughter, hesitation and self-correction to speech and starts playback before a sentence is finished.
Music and sound for media
Gen mixes speech, sound effects, ambience and background music in one clip, and Music turns lyrics or dry vocals into full songs.
What's included with StepAudio 3?
Expressive streaming speech, context-aware transcription, a full-duplex voice agent, voice design and cloning, full-scene audio and song generation.
- Expressive streaming speech
- TTS adds laughter, hesitation, stutters and self-correction, and streams audio so playback can start before a sentence is done.
- Context-aware transcription
- ASR Max uses the surrounding context to get names and jargon right, and copes with whispers, fast speech, singing and background music.
- Full-duplex voice agent
- Realtime lets both sides talk at once, stops when it is really interrupted, thinks while it speaks and runs tools without pausing the talk.
- Voice design and cloning
- Gen builds a voice from a written description, and TTS can use a cloned voice for $1.50 per voice.
- Full-scene audio
- Gen puts several speakers, sound effects, ambience and background music into one clip from a single prompt.
- Songs from lyrics or a vocal
- Music writes songs, instrumentals and covers up to 5 minutes 30 seconds long, and can build an arrangement around an a cappella take.
How much does StepAudio 3 cost?
StepAudio 3 is paid per use on StepFun's API: StepAudio 3 TTS costs $0.36 per 10,000 characters, StepAudio 3 ASR Max $0.40 per hour of audio, and voice cloning $1.50 per voice, while Realtime, Gen and Music are free during their preview.
- Streams speech while it generates
- Up to 1,000 characters per request
- Six languages, four of them in preview
- Uses context to get names and technical terms right
- Handles dialects, whispers, singing and background music
- Text streamed back over HTTP and SSE
- Works with StepAudio 3 TTS and older StepFun TTS models
- Test calls pay only for the speech they make
- Charged once a production clone succeeds
StepAudio 3 Realtime, Chat, Gen and Music are free for a limited time as previews, and StepFun will replace them with paid versions. Two English letters count as one billed character, and one Chinese character counts as one. You pay from a prepaid balance, and rate limits rise with the total you have topped up, from 5 concurrent requests under $15 to 130 at $1,500 or more. Individual accounts must pass identity verification first. StepFun's Step Plan subscriptions do not cover the StepAudio 3 models. Full pricing on platform.stepfun.ai. We checked these prices on .
How do readers rate StepAudio 3?
No one has rated StepAudio 3 yet. Be the first.
Sign in to rate. One rating per account.
StepAudio 3 pros and cons
It is cheap and scores well in StepFun's own tests, but most of its languages are in preview and three of its five models have no final price yet.
Pros
- Cheap speech: StepAudio 3 TTS costs $0.036 per 1,000 characters, less than half of Eleven v3's $0.08.
- Scores near the top of StepFun's published voice-agent tables: 98.9 on the Artificial Analysis Full-Duplex Bench, against 95.3 for GPT-Realtime-2.
- ASR Max posts low error rates in StepFun's tests, 1.18% on LibriSpeech clean and 0.49% on AISHELL-1, and it can transcribe singing.
- Realtime, Gen and Music cost nothing while they are in preview.
Cons
- Narrow language support: Realtime speaks only Chinese and English, and four of the six TTS and ASR languages are still in preview.
- Three of the five models are previews with no published price, and StepFun says the preview versions will be retired, so model IDs and costs will change.
- Individual accounts need identity verification, and new accounts get 5 concurrent requests until they top up $15 or more.
- Early users of StepFun's Music demo on Hugging Face reported song-cover uploads timing out, and the weights have not been released.
What are the best StepAudio 3 alternatives?
GPT-Realtime-2 for voice agents at a fixed price, Eleven v3 for speech in 70+ languages, and Scribe v2 for cheaper transcription in 90+ languages.
GPT-Realtime-2PaidChoose it when you need a voice agent at a fixed, published price today; it costs $32 / $64 per million audio input / output tokens, while StepAudio 3 Realtime is a preview StepFun will retire.
Eleven v3FreeChoose it when you need speech in more than six languages; it covers 70+ and has a free plan, though it costs $0.08 per 1,000 characters against $0.036 for StepAudio 3 TTS.
Scribe v2FreeChoose it when you transcribe many languages or want a lower price; it covers 90+ languages at $0.22 an hour, against $0.40 for ASR Max.
Questions about StepAudio 3
What is StepAudio 3?
StepAudio 3 is StepFun's family of five audio models, for speech, transcription, live voice agents, sound and music, sold per use on its API to developers, especially those working in Chinese and English.
Who is StepAudio 3 for?
StepAudio 3 is best for developers who want speech synthesis, transcription, real-time voice agents and music generation from one API, especially for Chinese and English audio.
Is StepAudio 3 free?
Partly. StepAudio 3 Realtime, Chat, Gen and Music are free on StepFun's API for a limited time while in preview, and Voice Studio lets you try speech for free. StepAudio 3 TTS and ASR Max are billed per use from a prepaid balance.
How much does StepAudio 3 cost?
StepAudio 3's paid plans are StepAudio 3 TTS at $0.36 per 10,000 characters, StepAudio 3 ASR Max at $0.40 per hour of audio and Voice cloning at $1.50 per voice.
What are the best StepAudio 3 alternatives?
The closest alternatives on Recaply are GPT-Realtime-2, Eleven v3 and Scribe v2.
Where can I try StepAudio 3?
On its official website, static.stepfun.com.
Which languages does StepAudio 3 support?
StepAudio 3 TTS and ASR handle Chinese, English, Japanese, Korean, French and Spanish, though every language except Chinese and English is still in preview. StepAudio 3 Realtime works in Chinese and English only.
Is StepAudio 3 included in StepFun's Step Plan?
No. StepFun's Step Plan subscriptions, from $6.99 to $99 a month, cover its text models plus StepAudio 2.5 TTS and ASR, but not the StepAudio 3 models. StepAudio 3 is billed per use from your prepaid StepFun API balance.
Is StepAudio 3 open source?
No. StepFun offers StepAudio 3 only through its API and Voice Studio, and has published technical reports but no model weights. Some earlier StepFun audio models, such as Step-Audio 2 mini, were released openly on Hugging Face.
How long can StepAudio 3 Music songs be?
Up to 5 minutes 30 seconds, according to StepFun's technical report, in 48 kHz stereo. It writes songs with vocals, instrumentals and covers from a style prompt, lyrics or a dry vocal recording.
Our verdict
StepAudio 3 is a strong low-cost pick for Chinese and English voice work, and a gamble for anything else.
Its TTS costs less than half of Eleven v3, and StepFun's own tests put Realtime and ASR Max at or near the top, but three of the five models are unpriced previews.
- Skip it for
- Products that need many languages, or a voice agent with a production price today.
- Check first
- Check that the languages you need are out of preview, and budget for Realtime, Gen and Music once they turn paid.
Step-Audio-R1
Voxtral 3
Fish Audio S2.1 Pro
SAM Audio
Scribe v2 Realtime
Gemini 3.5 Transcribe