Voice Arena 发布 Dictation Bench：首个面向真实听写应用的开放基准
团队把 761 段录音播给 10 款听写应用实测，Wispr Flow 拿到 79% 可直接发送率，Apple 只有 35%。评测按编程、邮件等 7 个场景分开展开，还有分场景细节。如果你日常靠语音输入写代码或回消息，这份榜单能帮你直接选工具。
Shobhit Banga 团队发布了 Voice Arena Dictation Bench，据称是首个针对 AI 听写应用（而非 API）的开放基准。测试方法是将 87 名说话人的 761 段录音，在同一台 Mac 上逐一播放给 10 款听写应用，再由 5 个 LLM 评审判断每份草稿能否不经修改直接发送。结果差距显著：Wispr Flow 以 79% 的可用率排第一，Apple 的听写仅 35%。
该基准的方法设计与常见语音评测不同：它按 7 个使用场景分类（如编程工作流、邮件消息等），而不是用一段通用文本测所有工具；同时把“需要修正的草稿频率”和“字符编辑总量”两个指标分开统计，并展示具体例子说明差异。核心评判标准是语义保真——文本流畅但改动用户想传达的关键细节（如技术术语）也算失败。
对开发者和重度听写用户的实际意义在于：可以先用基准数据缩小候选范围，再用自己的语音实测延迟、隐私和工作流契合度，减少对产品演示的依赖。
A dictation benchmark should help someone choose a tool, not require them to become a speech researcher first. The useful starting question is what they need the tool to write. For a coding workflow, exact technical terms matter. For messages and emails, preserving the intended meaning through restarts and corrections matters. A product can produce fluent text and still fail the task if it changes the detail the user needed to communicate. @voicearena_ai 's Dictation Bench is organized around seven use-case categories rather than one generic passage. It also separates the frequency of drafts needing a correction from the total volume of character edits, and shows examples that make those differences understandable. If I were choosing a tool, I would use that evidence to build a shortlist, then test it on my own speech. Latency, privacy, and workflow fit would still matter, but I would start with a clearer picture of output quality. That is the practical significance of a thorough evaluation: it makes a real decision less dependent on a polished demo. The research matters because it helps answer a question people already need to resolve. Shobhit Banga @shobhitbanga Drop 4/15: Announcing the Voice Arena Dictation Bench - the first open benchmark of AI dictation apps. Not their APIs: the real apps. We played 761 recordings from 87 speakers into each of 10 apps on a Mac, then 5 LLM judges asked if each draft could be sent as-is. Wispr Flow: 79%. Apple: 35% 🧵
This report is categorized under AITOP's "Products" desk, focusing on US frontier AI labs and Silicon Valley technology leaders. Within our cross-regional analytical framework, it directly maps to AI developer toolchains, coding agents, and engineering workflows and US-China foundation model benchmarks and commercialization comparison. Key entities and technical topics under continuous tracking include Wispr Flow, 语音识别, 听写, Voice Arena.
Source attribution: curated from elvis (@omarsar0).
