What is StepAudio 3 TTS?
StepAudio 3 TTS is the text-to-speech member of StepFun's StepAudio 3 family, five audio models released on 15 September 2026 alongside a full-duplex conversation model, speech recognition, general audio generation and music. It streams speech as it generates and treats laughter, breaths, hesitation and self-correction as part of a performance rather than noise to be removed.
There are twelve built-in voices, seven carried over from StepAudio 2.5 and five new. Delivery is set with a plain-language instruction of up to 500 characters, or with a direction in parentheses inside the text, which is acted on but never read aloud. Voices can be cloned from reference audio or designed from a description. English and Chinese are fully supported; Japanese, Korean, French and Spanish are in preview.
The model is API-only through StepFun's Chinese and international platforms and partners such as EmpirioLabs. No weights have been published, unlike StepFun's older Step-Audio-TTS-3B.
How much does StepAudio 3 TTS cost?
| Plan | Price | What you get |
|---|---|---|
| Speech | $14 per 1M billable characters | English letters, digits and spaces count 1; Chinese 3; Japanese 2.5 |
| Voice cloning or design | $1.50 per voice | Charged when the voice is added to the library |
$14 per 1M billable characters from 23 November 2026 (launch price was reported at $0.36 per 10,000); voice cloning or design $1.50 per voice. Chinese characters count as 3, Japanese as 2.5.
Is StepAudio 3 TTS worth it?
StepAudio 3 TTS is the reason the top of Artificial Analysis's speech arena is no longer an all-Western affair. It sits third, with 1,307 Elo, 33 points behind ElevenLabs' Eleven v4 and ahead of Cartesia's Sonic 3.6, at $14 per million characters, against $80 for Eleven v4 and $49 for Sonic 3.6 in the same table. Few speech models sound this good for this little.
The design reflects StepFun's view that hesitation is part of how people speak. You tell it how to perform in a sentence, or tuck a stage direction in brackets, and it adds the breaths and false starts that most systems sand off. That suits drama, explainers and character voices more than a bank's hold message.
The catch: the plumbing is narrow. Each request is capped at 1,000 characters, so a chapter has to be cut into sentence-sized pieces and stitched back together, and output tops out at 24 kHz. Four of its six languages are previews. Billing counts a Chinese character as three and a Japanese one as 2.5, so the headline price flatters English. And the price itself has moved: StepFun's pricing page changes rates on 23 November 2026, and a $0.36 per 10,000 figure from launch still circulates on third-party trackers.
Our take: for English or Mandarin narration on a budget, this is now the voice to beat. Teams that need long single requests, high sample rates or broad language coverage will spend the savings on engineering.
Where it earns its place
A two-person studio making English-language audio drama
Voice a 40,000-word script with breaths, pauses and hesitations that sound unscripted
Split the script into requests under 1,000 characters, add bracketed directions per line and stitch the audio. About 240,000 characters cost roughly $3.40 at $14 per million.
A Shanghai e-learning company
Narrate Mandarin lessons with a consistent cloned teacher voice
Clone the teacher once for $1.50, then synthesise lessons; budget for each Chinese character counting as 3 billable characters, about $42 per million characters of script.
Before you commit
- Setup
- POST to an OpenAI-style /v1/audio/speech endpoint with model stepaudio-3-tts, a voice and optional instruction; non-streaming responses return an audio URL.
- Integrations
- StepFun international and Chinese platforms; resold by EmpirioLabs; mp3, wav, flac, opus and pcm output at 8 to 24 kHz.
- Data and privacy
- API only, no weights. StepFun's pricing page and the docs we read do not state retention or training use for audio and text; check the platform terms before sending customer content.
- Cost at scale
- 10 million characters of English narration cost $140; the same length of Chinese text bills as 30 million characters, $420.
- Limits
- 1,000 characters per request; 48 kHz rejected. New accounts start at 5 concurrent requests and 100 requests a minute, rising with total top-ups.
What’s good
- Third on Artificial Analysis's arena at 1,307 Elo
- $14 per 1M characters, under a fifth of Eleven v4's $80
- Plain-language direction and natural breaths and hesitations
- Cloning or designing a voice costs $1.50
What’s not
- 1,000 characters per request, 24 kHz maximum
- Japanese, Korean, French and Spanish still in preview
- Chinese characters bill as 3, Japanese as 2.5
- Data-handling terms for audio not stated in the docs we read
Best for
Producers of English or Mandarin narration, explainers and character dialogue who want near-top quality at a low per-character price and can split scripts into short requests.
StepAudio 3 TTS alternatives
StepAudio 3 TTS: common questions
- Is StepAudio 3 TTS worth it?
- Near the top of the listener rankings for a fraction of the leaders' price, provided your scripts are short and mostly in English or Chinese. We rate it 4.2 out of 5.
- How much does StepAudio 3 TTS cost?
- $14 per 1M billable characters from 23 November 2026 (launch price was reported at $0.36 per 10,000); voice cloning or design $1.50 per voice. Chinese characters count as 3, Japanese as 2.5.
- What are the downsides of StepAudio 3 TTS?
- 1,000 characters per request, 24 kHz maximum; japanese, Korean, French and Spanish still in preview; chinese characters bill as 3, Japanese as 2.5; data-handling terms for audio not stated in the docs we read.
- Who is StepAudio 3 TTS best for?
- Producers of English or Mandarin narration, explainers and character dialogue who want near-top quality at a low per-character price and can split scripts into short requests.
- How much does StepAudio 3 TTS cost for a team or in production?
- 10 million characters of English narration cost $140; the same length of Chinese text bills as 30 million characters, $420.
- Is StepAudio 3 TTS safe for company data?
- API only, no weights. StepFun's pricing page and the docs we read do not state retention or training use for audio and text; check the platform terms before sending customer content.
- What are the best StepAudio 3 TTS alternatives?
- The closest alternatives we have reviewed are Eleven v4 and v4 Turbo, Cartesia Sonic 3.6, Qwen-Audio-3.1-TTS-Plus and Speechify Simba 3.5.
Answers checked Nov 21, 2026.