Wispr Flow releases Canto, a speech model trained for noisy real-world dictation

A dictation model aimed at everyday audio

Wispr Flow has announced Canto, a speech recognition model built for real-time dictation as people actually do it: speaking into laptop microphones and earbuds from desks, commutes and busy offices, with other voices, music or traffic in the background. According to the company's announcement, which reached the front page of Hacker News, Canto posted the lowest word error rate among all the models Wispr tested on an evaluation of real-world dictations, ahead of systems from Google, OpenAI, AssemblyAI and Deepgram. It is the first model out of the Wispr Advanced Interfaces Lab, which the company frames as the start of a broader research program.

How the benchmark was built

To measure performance in realistic conditions, Wispr assembled an evaluation set of 10 hours of English-language dictations from its own product, drawn from more than 2,300 unique speakers and sampled at random across applications and use cases. The company says it kept training and test speakers strictly separate to avoid overfitting to particular voices, and that every sample came from a user who had opted in to its data-sharing setting. On that set, Canto achieved the lowest word error rate — substitutions, insertions and omissions relative to a human transcript — of the compared models. Worth noting: the evaluation is Wispr's own, run on data from its product's usage distribution.

The stress test

Random samples were not the whole story. Wispr also built a separate three-hour challenge set containing the conditions most likely to break dictation: nearby speech, music, traffic, wind, low recording volume, whispered or far-field speech, and very short utterances with almost no surrounding context. On the full challenge set, Canto ranked second overall, behind Gemini 3.1 Pro, a frontier-size multimodal model that Wispr notes is too large for real-time, low-latency applications. Among the real-time models evaluated, Canto still had the lowest error rate. Broken down further, Gemini 3.1 Pro did best on noisy audio, while Canto tied for best on low-volume speech and on short dictations. Clips of only one or two words produced the highest error rates for every model in the comparison, since a single mistake weighs heavily on a short utterance and there is little linguistic context to resolve ambiguity.

Public benchmarks tell a different story

On three public English datasets, Canto tied for the lowest error rate on LibriSpeech but did not lead on FLEURS or Common Voice, where Wispr describes its results as competitive. The company argues these datasets consist mostly of read speech — audiobooks and prepared sentences — a different distribution from spontaneous dictation, where speakers pause, revise mid-thought and provide minimal context. That gap is Wispr's stated reason for evaluating on both public benchmarks and its own usage data.

Supervised fine-tuning plus reinforcement learning

Canto starts from a model pretrained on millions of hours of speech and text, then trains in two stages. Supervised fine-tuning teaches the transcription task by pairing audio with reference transcripts. A reinforcement learning stage then has the model generate several candidate transcripts for the same audio; each candidate is scored against a reference, and those scores steer future training toward better outputs. Wispr uses Group Relative Policy Optimization (GRPO), the method introduced in DeepSeekMath, comparing candidates within each group and learning from their relative rewards. Sequence-level training of this kind has a long history in speech recognition, including minimum word error rate training, but direct applications of GRPO to autoregressive speech recognition are recent. Wispr says the infrastructure it built to generate and score speech rollouts at scale also supports research into contextual recognition, personalization, diarization and difficult audio. The company is also working on learning from user corrections, since vocabulary shifts faster than models can be retrained — its example is a word like "Claude" that once read as "cloud".

Why it matters

Most speech recognition benchmarks measure clean, read audio, while most real dictation happens in noisy rooms through mediocre microphones with almost no context. Canto stands out mainly because Wispr is explicitly optimizing and measuring for those real conditions, and reporting both where it wins (its own data distribution) and where it does not (behind a frontier model on hard audio, and not leading on public read-speech datasets). The GRPO-based recipe — targeting specific failure modes with reinforcement learning rather than chasing aggregate accuracy alone — is a pattern other speech teams will likely watch as voice becomes a primary input modality. The claims remain vendor-run, so independent verification will be the real test.