WisprFlow trains its own Canto 1 speech model
The shift from clean-audio benchmarks to messy laptop-mic conditions changes what buyers should demand when evaluating dictation tools. Developers building voice interfaces can borrow Wispr's data-labeling trick of separating genuine mishearings from stylistic rewrites. Enterprise teams comparing transcription vendors get a concrete accuracy figure measured on real usage rather than read-aloud corpora.
WisprFlow has moved from rented speech engines to models it trains in-house and hired Ariya Rastrow, the founding Alexa scientist, as chief science officer. Its first model, Canto 1, is built to reduce how often people have to edit dictated text. On about 10 hours of real Flow dictations from more than 2,300 speakers, Canto 1 posted a 3.4% word error rate, lower than all competing models Wispr tested.
The model was trained on noisy real-world dictations rather than the clean, read-aloud audio that speech recognition was historically benchmarked on, and it is tuned to each user's vocabulary and names. Since launch, dictations that take more than one second to appear have dropped by 54%, which matters because typing gives feedback on every keystroke while speech forces users to wait until a sentence ends.
Training follows the same stages used for large language models: pretraining, supervised fine-tuning, and reinforcement learning with GRPO. The RL stage uses opt-in corrections from Flow users, where a misheard word aligned with audio is folded back into training references while plain rewrites are ignored.
The keyboard was a workaround for machines that couldn’t hear. Machines that hear make it optional.
WisprFlow recently hired Ariya Rastrow (founding Alexa scientist) as chief science officer and moved from rented speech engines to models it trains itself.
Hard to find a better fit for a speech lab than someone who scaled Alexa’s speech team.
Canto 1, his lab’s first model, was built to cut how often people edit dictated text, so Wispr trained it on noisy real-world dictations and tuned it to each user’s vocabulary and names.
And this is a hard test, unlike many, because Speech recognition was long tested on clean, read-aloud audio, but the hard problem is a cheap laptop mic in a room where someone else is talking.
On about 10 hours of real Flow dictations from more than 2,300 speakers, it posted a 3.4% word error rate, lower than all competing models that Wispr tested.
Typing gives you feedback on every keystroke, while speech makes you wait until the end of a sentence to see what the machine heard, so speed counts as much as accuracy.
Since launch, dictations that take more than 1 second to appear are down 54%.
Its training runs through pretraining, supervised fine-tuning and reinforcement learning with GRPO, the same stages used for large language models.
The reinforcement learning stage uses corrections from Flow users who opted in to share data.
When someone fixes a dictated word, Wispr checks how the audio aligns with the text and how the edit is shaped.
A misheard word gets folded back into the training reference, and a plain rewrite of the sentence gets ignored.
This report is categorized under AITOP's "Models" desk, focusing on US frontier AI labs and Silicon Valley technology leaders. Within our cross-regional analytical framework, it directly maps to AI developer toolchains, coding agents, and engineering workflows and US-China foundation model benchmarks and commercialization comparison. Key entities and technical topics under continuous tracking include WisprFlow, Canto 1, speech recognition, GRPO.
Source attribution: curated from rohanpaul_ai (@rohanpaul_ai).
