Step-Audio TTS 3B
Step-Audio TTS 3B uses a dual-codebook language-model design for speech synthesis. The repository includes a matching vocoder and a separate vocoder optimized for humming.
Use the correct waveform component
The release separates the language model's audio-token generation from waveform reconstruction. Humming has its own optimized vocoder.StepFun Step-Audio-TTS-3B model card
Keep speech and musical evaluation separate
Practical guidance. A low transcription error does not measure melody or timing. Evaluate humming and rap with listening tasks that include pitch, rhythm and continuity, then assess spoken content separately.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Access | Downloadable |
| Model identifiers | Step-Audio-TTS-3B |
| License | Apache 2.0.StepFun Step-Audio-TTS-3B model card |
| Training approach | The card describes a large synthetic dataset and an LLM chat training format.StepFun Step-Audio-TTS-3B model card |
| Output range | Speech, rap and humming, with language, emotion and style control.StepFun Step-Audio-TTS-3B model card |
Related models
- Step-Audio Chat
StepFun · Step-Audio
Step-Audio Chat is the 130B multimodal language component of Step-Audio.
Launch date not confirmedDownloadable - Step-Audio EditX
StepFun · Step-Audio EditX
Step-Audio EditX is a 3B audio model trained with reinforcement learning for iterative changes to emotion, speaking style and vocal effects.
Launch date not confirmedDownloadable - StepAudio 3 TTS
StepFun · StepAudio 3
StepAudio 3 TTS produces speech from text and reference audio.
Launch date not confirmedAPI · Provider product