Canto for everyday dictation

A dictation model is judged less by perfect studio speech than by the one-line message recorded next to traffic, another speaker or a laptop fan. Wispr says Canto was tuned for this kind of spontaneous input. Its release study used ten hours of opt-in English dictations from more than 2,300 speakers, sampled across apps and use cases, with train-test speaker separation. Wispr reports the lowest word error rate among the models in that comparison, but those internal recordings and selections are not an independent benchmark.

The vendor also assembled a three-hour challenge set of noisy, low-volume, whispered, far-field and short dictations. Canto ranked second overall behind Gemini 3.1 Pro on that set and was the strongest of the real-time transcription models Wispr evaluated. This nuance matters: the best overall model in an offline comparison may be too slow or expensive for live writing. Short one- or two-word utterances remained difficult across models. Teams should test proper nouns, jargon, accents and noisy devices they actually use rather than assume the headline ranking transfers unchanged.

Canto is part of the Wispr Flow experience, which can dictate into supported applications and use a personal dictionary for specialized terms. Wispr’s data-controls page says transcription is cloud-based even when users opt out of model improvement. Audio, transcripts and edits may be used for evaluation or training when the user enables sharing; enterprise accounts have sharing disabled. Buyers handling confidential speech should examine these settings, retention and contractual protections before adoption. Wispr’s published plans price the Flow app by dictation allowance and administration, not by a standalone Canto model invocation.