Canto is a 📝Wispr Flow speech recognition model built for real-time dictation, converting speech to text under the conditions people actually dictate in rather than clean, controlled ones.
It was introduced on September 17, 2026 as the first public release from 📝Wispr Advanced Interfaces Lab, starting from a model pretrained on millions of hours of speech and text and then post-trained through supervised fine-tuning followed by reinforcement learning. On ten hours of real Flow dictations sampled from more than 2,300 opted-in speakers, Wispr reports it achieved the lowest word error rate of every model tested, including systems from 📝Google, 📝OpenAI, AssemblyAI, and Deepgram; on a harder three-hour challenge set it placed second to 📝Gemini 3.1 Pro, a frontier-scale multimodal model unsuited to low-latency work, and first among real-time transcription models.
Canto is not sold or served on its own — it reaches people as the recognition engine inside Flow, where it draws on each user's personal dictionary at dictation time. Published evaluation is English-only so far, with matching quality across more languages named as a goal for the successor model already training at more than ten times Canto's scale.
Key Features
- Real-world accuracy — Evaluated on ten hours of everyday Flow dictations from more than 2,300 speakers, where it recorded the lowest word error rate of any model Wispr tested.
- Difficult-audio handling — Holds up against nearby speech, music, traffic, wind, whispering, and far-field audio, tying for the lowest error rate on low-volume and short samples.
- Contextual vocabulary — Reads names and rare terms from a user's personal dictionary at runtime, trained against phonetic distractors so it follows the audio when the two disagree.
- Correction grafting — Uses alignment confidence and the shape of an edit to separate genuine mistranscriptions from rewrites, folding only the real fixes into its training targets.
- GRPO post-training — Generates several candidate transcripts per clip and scores them against each other, letting Wispr train directly against named real-world failure modes.
Getting Started
- Install Wispr Flow on Mac, Windows, or Android and sign in to the app.
- Dictate with the Flow hotkey in any application; Canto transcribes in real time.
- Add names, jargon, and acronyms to your Flow dictionary for Canto to draw on.
- Opt into data sharing if you want your corrections to inform future models.