Wispr introduced the Canto speech recognition model

Company Wispr released the Canto model designed for high-precision real-time speech recognition. Developers position the solution as a tool for working in complex acoustic conditions where minimal delay and high resistance to background noise are required.

In internal tests on a 10-hour dataset simulating real dictation scenarios, Canto demonstrated the lowest Word Error Rate (WER) among all available real-time models. According to the company, the architecture surpassed in accuracy similar solutions from Google, OpenAI, AssemblyAI, and Deepgram.

The model is optimized for operation in conditions where standard recognition systems often make mistakes due to accents, specific terminology, or low recording quality. Technical details of the architecture and testing methodology are presented on the official project page.

The release of Canto is aimed at bridging the gap between laboratory benchmarks and practical application of speech recognition technologies. Unlike many universal models, this solution is focused on minimizing delays while maintaining transcription accuracy in dynamic environments.