Step-Audio

What it is Step-Audio is a set of open speech-interaction models from StepFun (阶跃星辰). Speech interaction means talking with a model by voice, with the model both understanding spoken input and producing spoken output. The release on 17 February 2025 included three components: Step-Audio-Chat, with 130 billion parameters, Step-Audio-TTS-3B and Step-Audio-Tokenizer ^[1].

Origins StepFun published the Step-Audio technical report on 17 February 2025 and released its inference code and weights at the same time. The technical report is arXiv:2502.11946 ^[2]. The release was a joint open-sourcing with Geely, alongside Step-Video-T2V ^[3].

Main features The models support Chinese, English and Japanese, as well as dialects including Cantonese and Sichuanese. They can control emotions such as anger, joy and sadness, and vocal styles including rap and singing ^[1]. They also support voice cloning, which means producing speech in a particular person's voice, and tool-calling agent features, which let the model call external functions while it converses ^[1].

Hardware needs are high. Step-Audio-Chat needs at least 265GB of GPU memory. StepFun recommends four A800 or H800 GPUs with 80GB each ^[1].

Step-Audio-TTS-3B reports a character error rate of 1.31 on the SEED test-zh set and a word error rate of 2.31 on test-en ^[1]. Character and word error rates measure how often a transcription differs from the reference text, so lower numbers are better.

StepFun's StepEval-Audio-360 benchmark was judged by GPT-4o. On it, Step-Audio-Chat scored 4.11 for chat, against 3.49 for GLM4-Voice and 2.27 for Qwen2-Audio ^[1]. These are StepFun's results on its own benchmark.

Licence and status The code is released under Apache 2.0. The weights are under separate terms in the Hugging Face repositories ^[1].

The repository is no longer maintained. StepFun points users to Step-Audio 2, Step-Audio-R1 and Step-Audio-EditX as its successors in this area ^[1].

Place in the field Step-Audio was among StepFun's first open speech models. Its technical report and benchmark show a focus on speech and dialogue, and newer projects from the same company have since replaced the open repository.