StepFun Releases Five StepAudio 3 Models Covering Full Audio AI Stack

On September 15, StepFun released the StepAudio 3 family—five models spanning TTS, Gen, Realtime, ASR and Music—covering listening, speaking, understanding, interaction and creation, with full API documentation going live simultaneously. The release signals that audio AI competition is shifting from single-point models to full-stack systems.

Confirmed

• The realtime model is organized around a continuous listen-speak-think-act loop, with Deep Perception and Seamless Duplex capabilities, supporting interruption, reasoning while speaking, and tool calls
• The ASR model achieves 1.7% WER; the realtime model scores 98.9% on Artificial Analysis' conversation naturalness and tops that leaderboard
• StepAudio 3 Music supports explicit music planning (including ABC notation) and open-domain text-controlled generation up to 5.5 minutes; its Music Tokenizer represents audio as a single-codebook stream at 50Hz with a 65536 codebook, using a flow-matching architecture
• API docs list the full lineup including stepaudio-3-tts (low-latency streaming TTS), stepaudio-3-asr-max and stepaudio-3-realtime

Why it matters

• New智元 (new media outlet) commented that launching five models at once shows audio AI competition has escalated to full-chain systemic rivalry
• The realtime model emphasizes voice-agent capabilities—interruption, thinking while speaking, tool invocation—targeting realtime interaction rather than mere transcription or synthesis

2026-09-15 ~ 2026-09-16 · 6 related posts

• [source] StepFun drops five StepAudio3 models, claiming three Artificial Analysis world No.1s — 新智元 · 2026-09-15
• StepFun launches StepAudio 3: five audio models, #1 in realtime speech evals — StepFun_ai · 2026-09-16
• [source] StepAudio 3 API docs detail full model lineup from TTS to music generation — StepFun_ai · 2026-09-16
• StepFun's StepAudio 3 Realtime reasons while speaking, hits 90.6 on MMSU — stepfun-ai · 2026-09-16
• StepAudio 3 Music plans songs with ABC notation, ranks #3 behind Suno V5.5 — stepfun-ai · 2026-09-16

1 near-duplicate retellings: StepFun_ai