Wispr Flow 发布 Canto：实时听写 WER 3.4%
Wispr Flow（跨端 AI 听写）上线自研语音模型 Canto。在 9.8 小时、2300+ 说话人的英语 Flow 听写评测上达到 3.4% 词错误率，为其所测最低；面向嘈杂/口音等真实场景，Series B 时已预告难点错误率可从 30%+ 降至约 5–10%，减少事后改稿。
Introducing Canto, our latest speech model for real-time dictation. On our evaluation of 9.8 hours of English Flow dictations from more than 2,300 speakers, Canto achieved a 3.4% word error rate—the lowest among the models we tested. Learn more about Canto from our CSO, Ariya Rastrow: Research notes: - Product: https://wisprflow.ai — AI voice-to-text for polished writing in every app (Mac/Windows/iPhone/Android); Notetaker also available - Canto: proprietary speech model for Flow real-time dictation; eval WER 3.4% on 9.8h English Flow dictations / 2,300+ speakers - Series B context (2026-08-17): https://wisprflow.ai/post/series-b — Canto preview; hardest conditions WER from 30%+ down to ~5–10%; expect ~30–35% fewer dictations needing edits; code-switching / personal vocab - Official @WisprFlow; CSO Ariya Rastrow; posted 2026-09-17