StepAudio 3 ASR 的领先来自面向智能体用例的对话音频——在 AA-AgentTalk 上（占 AA-WER Index 一半权重），它取得 1.4% WER，领先 ElevenLabs Scribe v2 的 1.5% 和 Fun-Realtime-ASR-preview 的 2.0%；在 VoxPopuli-AA 上，它与 Fun-Realtime-ASR-preview 持平，均为 1.2%。在 Earnings22-AA（长篇财报电话会议）上，它得分 2.8%，而 Fun-Realtime-ASR-preview 为 1.8%，这是其三个数据集中最弱的一项。
StepAudio 3 ASR's lead comes from conversational audio focused on agentic use cases - on AA-AgentTalk, which carries half the weight of the AA-WER Index, it scores 1.4% WER, ahead of ElevenLabs Scribe v2 at 1.5% and Fun-Realtime-ASR-preview at 2.0%, and on VoxPopuli-AA it matches Fun-Realtime-ASR-preview at 1.2%. On Earnings22-AA, long-form earnings calls, it scores 2.8% against Fun-Realtime-ASR-preview's 1.8%, the weakest of its three datasets.
来源：@ArtificialAnlys · x.com