Wispr's self-trained Canto 1 speech model hits 3.4% WER on real-world dictation
rohanpaul_ai · x · 2026-11-16
Wispr Flow hired founding Alexa scientist Ariya Rastrow as chief science officer and moved from licensed speech engines to in-house models. Its first model, Canto 1, targets the hard case — noisy laptop-mic dictation: on 10 hours of real Flow dictations from 2,300+ speakers it hit a 3.4% word error rate, lower than all competitors tested; dictations taking over 1 second to appear are down 54% since launch. Training follows LLM-style stages — pretraining, SFT, and GRPO-based RL on opt-in user corrections, folding only misheard-word fixes into training references while ignoring full rewrites. A Voice Arena chart shows Wispr Flow needs edits on just 20.6% of drafts vs Gemini 24.2%, ChatGPT 35.7%, Apple Dictation 65.4%.
Related event: Wispr Flow Launches In-House Voice Model Canto 1(2 posts)→
More from Models
- Anthropic tightens supported regions policy, closing loopholes for restricted countries' entities — J_MaestreVidal · 2026-11-16
- In real-world testing, Gemini 4 ranks a distant third for coding behind Claude Fable 5.2 and GPT-6 Astra — HarrySurden · 2026-11-16
- Why no GPT-6.2 yet? OpenAI may be prioritizing scarce compute for research over shipping intermediate checkpoints — kimmonismus · 2026-11-16
- AI influencer teases: November to be a good month for open models — mervenoyann · 2026-11-16
- inside-llm: open-source tool visualizes LLM internals and token-by-token quantization diffs — AlternativeEnough352 · 2026-11-16
- repligate: Fable 5 models get cut off just for thinking about classifiers — repligate · 2026-11-16