Canto beats Google and OpenAI on real-world dictation
Canto: A speech model built for the real world
Wispr Flow's Advanced Interfaces Lab introduces Canto, a speech model trained for real-time dictation in noisy, everyday conditions. On 10 hours of real user dictations, Canto achieved the lowest word error rate among models from Google, OpenAI, AssemblyAI, and Deepgram. It ranked second only to Gemini 3.1 Pro on a brutal challenge set, but first among real-time models. The team used supervised fine-tuning and reinforcement learning with GRPO, plus a clever technique to learn from user corrections.
We built infrastructure that can generate and score speech rollouts at scale, allowing us to construct training environments around specific behaviors and failure modes.
- ks2048
They need to show some examples. You beat all the top models on your private data set? Show at least a couple examples - audio and transcripts - from examples that your model got right and others got wrong.
- Redster
Congrats on the launch! I'm glad more progress is being made in this area.
Because of the hallucinations inherent in transformer models, I went looking for a transducer-based model with a low WER. I have been super pleased with parakeet-unified-en-0.6b. It's WER isn't as low as Canto, but it's about as low as you can get (~5-6.5%) with a non-transformer-based model as far as I'm aware.
I've been very pleased with its output.
I wasn't looking for this, but it's also lightweight enough to run on my little potato PC, which has an i5 8th gen processor, and still transcribe 9-10x faster than realtime.
I vibe-coded a little wrapper for it and use it on folders of audio or podcast rss feeds or even YT playlists and channels and it's been one of my new favorite tools.
- nr378
Is it actually better than Microsoft's MAI-Transcribe-2? That generally seems like the best model right now and it's not included in their benchmarks.
I switched from Superwhisper->WisprFlow->Spokenly->Fieldwork and found WisprFlow the least accurate of the 4.