Clean audio, messy reality

Speech models excel at transcribing controlled audio, but real dictation happens through laptop mics and earbuds in busy offices, cars, and cafes with background noise and competing voices.

Introducing Canto

The Wispr Advanced Interfaces Lab released Canto, a speech model for real-time dictation that achieved the lowest word error rate on real-world dictations compared to Google, OpenAI, AssemblyAI, and Deepgram.

“Today, the Wispr Advanced Interfaces Lab is introducing Canto, our latest speech model for real-time dictation.”

Testing on real dictations

Canto was evaluated on 10 hours of English dictations from over 2,300 unique speakers using Wispr Flow across different apps and use cases, with strict separation between training and test speakers to avoid overfitting.

Word error rate wins

Canto achieved the lowest WER, which measures word substitutions, omissions, and insertions against human-transcribed references. Lower WER is better.

Stress-testing with challenges

A separate 3-hour challenge set featured the toughest conditions: nearby speech, music, traffic, wind, low volume, whispered speech, and short dictations with minimal context.

Second place on hard audio

On the challenge set, Canto ranked second behind Gemini 3.1 Pro, a much larger frontier model unsuitable for real-time applications. Among real-time models, Canto had the lowest WER.

Short dictations are hardest

Short one or two word dictations produced the highest error rates across all models because a single mistake has outsized impact and models lack linguistic context to resolve ambiguity.

Public benchmarks tell partial story

Canto tied for lowest WER on LibriSpeech but remained merely competitive on FLEURS and Common Voice. These datasets contain read speech, not spontaneous dictation with pauses and revisions.

Two-stage training approach

Canto starts from a model pretrained on millions of hours of speech and text, then undergoes supervised fine-tuning with reference transcripts, followed by reinforcement learning to improve complete transcriptions.

Supervised fine-tuning explained

The model learns to predict words one step at a time by comparing its output against reference transcripts, teaching it the basic transcription task.

Reinforcement learning for outcomes

The model generates multiple possible transcriptions for the same audio, scores each against a reference, and learns to make better transcriptions more likely. This lets training focus on outcomes that matter, like handling difficult recording conditions.

Group Relative Policy Optimization

Canto uses GRPO, which compares candidate transcripts within each group and learns from their relative scores. Each candidate receives a reward measuring how well it meets the training objective.

Learning from user corrections

When users correct transcriptions, the model can learn. But not every edit signals an error. Wispr uses audio signals and edit patterns to identify genuine recognition mistakes versus rewrites or formatting changes.

Grafting corrections into training

Wispr selectively grafts likely corrections into the original transcript, creating a focused training signal. The model learns to make the intended correction and receives higher reward when it does.

Context without overconfidence

Canto uses personalized vocabulary to help recognize names and rare terms, but must learn when to trust context and when to trust audio. In poor conditions, both 'Barry' and 'berry' might sound plausible.

Training context discrimination

Wispr trained the model with phonetically similar distractor words alongside correct terms. This taught Canto to adopt correct context while resisting false suggestions that sound similar.

Bigger model coming

Canto is the first in a larger family. Wispr is training a successor at more than ten times Canto's scale to improve difficult audio, multi-speaker recognition, contextual vocabulary, and support more languages.

Unified transcription and diarization

Future models will recognize words and identify speakers simultaneously rather than in separate stages. Joint reasoning about what was said, who said it, and when speakers changed reduces compounding errors.

Beyond word recognition

Over time, Wispr's speech models will use context selectively, understand the application someone is speaking in, and connect what was said with what the person is trying to accomplish.