Overview
Dictation and transcription get treated as one job, and they are not. Transcription works from a recording, where a person reads a script, sentences are whole and the audio is clean. Dictation works from someone thinking: half a sentence, a change of mind, a name said wrong and said again, a pause to decide, another voice in the room. A model trained on clean recordings makes a specific kind of mistake in real dictation. It writes down the abandoned half-sentence too, because it genuinely heard it.
Dictation is what Wispr Flow does. The product sits on the cursor, so writing mail, drafting a document, typing a command in a terminal or prompting a model all use the same gesture. The company used other vendors' speech models until this year, when it set up the Wispr Advanced Interfaces Lab in July, brought in Ariya Rastrow, a founding scientist behind Alexa, as chief science officer, and shipped its first in-house model, Canto 1, on Sep 17. In November the independent Voice Arena dictation benchmark put Flow in first place.
That first place is worth reading properly. Voice Arena asks whether a draft could go out without touching it, and 79.4 percent of Flow drafts passed clean. A second measure counts the characters a person would have to change, where Flow scored 0.40 percent against 0.93 percent for the next best, more than double. Even when it was wrong, no draft needed more than 19 characters of fixing. Read together, those three numbers do not say every word is right. They say the errors are small, so you never rewrite a paragraph.
The gain from training in-house comes from a different training signal. Wispr does not use bought annotation. It uses what people actually change: for the same stretch of speech, the difference between what the model wrote and what the person kept. The hard part is telling two kinds of edit apart. One is mishearing. The other is the user changing their own mind, and learning from that would teach the model to treat someone's writing habits as recognition errors. Canto separates the two first and learns only from the first.
The method runs in two stages. Supervised fine-tuning pairs audio with reference transcripts so the model learns to predict each word. Then reinforcement learning has it generate several candidate transcripts for the same audio, scores each against the reference, and uses Group Relative Policy Optimization to make the better kind of result more likely. Supervision says what the word at each step should be. Reinforcement learning judges the finished document, which is closer to what actually matters: whether this can be sent.
Evaluation runs on two sets. One is 10 hours of English dictation sampled at random from more than 2,300 speakers across applications and use cases, with speakers in the test set strictly separated from training, and every sample from a user who opted into data sharing. The other is a 3 hour challenge set built from the conditions most likely to break dictation: nearby speech, music, traffic, wind, low recording volume, whispers, far-field audio, and very short dictations with almost no surrounding context.
The results need splitting. On the random sample Canto had the lowest word error rate of every model compared. On the challenge set it placed second, behind Gemini 3.1 Pro, which is a frontier-sized multimodal model that cannot run in a real-time low-latency setting, so Wispr adds that among real-time transcription models Canto had the lowest error rate. It tied for lowest on low-volume speech and on short dictations. On three public sets it tied for lowest on LibriSpeech and stayed competitive on FLEURS and Common Voice without leading either. Wispr says why: those sets are mostly read speech, a different distribution from everyday dictation.
Two product limits are worth knowing before switching. Notetaker is Mac and Windows only, so anyone who records meetings on a phone cannot use that part yet. And the free plan caps dictation by the week, 2,000 words on desktop and 1,000 on iPhone, unlimited on Android, a line that anyone writing long documents will hit quickly. Paid tiers run Pro, Growth and Enterprise. Pro gives unlimited dictation and unlimited Notetaker history. Growth is where single sign-on and a signed HIPAA agreement appear, and Enterprise adds automated provisioning, audit logs and device management. Teams with compliance requirements should not stop at Pro.
Whether it is worth using depends on what you want voice to be. As a faster keyboard for short messages, the free tier is enough. As your main way of writing, the number to watch is not accuracy but how often you edit, since every correction teaches you not to trust it. Wispr says Canto 2 is coming soon, and states the same test for it: whether people still have to fix what comes back.
Key Features
- Dictation on the cursor: Not tied to one app. Mail, documents, terminal commands and prompts for a model all use the same gesture, with the text landing wherever the cursor is.
- Trained on real edits: The training signal comes from what users change in a draft, with genuine mishearings separated from people changing their own mind, so writing habits are not learned as recognition errors.
- Two-stage training: Supervised fine-tuning learns the word at each position, then GRPO scores whole transcripts against a reference so the kind of draft that can be sent becomes more likely.
- An evaluation set built from real conditions: The challenge set collects nearby speech, music, traffic, wind, low volume, whispers, far-field audio and very short dictations, none of which appear in read-speech datasets.
- Restrained use of context: When a word sounds close to a name in the dictionary, it weighs the dictionary against what the audio supports instead of forcing the name through and rewriting what was clearly heard.
- Meeting notes priced separately from dictation: Notetaker captures audio on the device rather than joining as a visible bot, and currently runs on Mac and Windows only, with mobile on the roadmap.
Use Cases
- Writing mail and documents in a cafe or open office where other people are talking
- Dictating prompts into ChatGPT, Claude or Cursor instead of typing and retyping
- Replacing the keyboard for daily input for people with RSI or limited hand mobility
- Keeping product names and people's names consistent across a writing team with a shared dictionary
- Recording meetings on Mac or Windows and pulling up past transcripts afterwards
Pros
- Treats dictation as its own problem, optimising for whether a draft can be sent rather than for reproducing a recording word for word
- Uses real edits as the training signal and separates mishearing from a change of mind, which fits actual use better than piling up annotation
- Built a challenge set around nearby speech, low volume and far-field audio, so the evaluation is more credible than public dataset scores alone
- First on the independent Voice Arena benchmark with 79.4 percent of drafts untouched and 0.40 percent of characters changed, which means errors stay small
- Over 100 languages across four platforms, with Simplified Chinese by default and Traditional and Cantonese selectable
Pricing
Four tiers: Free, Pro, Growth and Enterprise. Free covers Mac, Windows, iPhone and Android with over 100 languages, dictionary and snippets, and limited Notetaker access. Pro adds unlimited dictation, unlimited Notetaker history, a shared dictionary and snippets, a team workspace, centralised billing and usage analytics, billed monthly or annually with the annual plan about 20 percent cheaper, priced in local currency by region. Growth adds single sign-on, a signed HIPAA agreement, multiple domains and an admin switch over model training. Enterprise is quoted, adding automated provisioning, audit logs, device management and domain capture.
Summary
Wispr Flow treats dictation as a separate problem, and its biggest move is training Canto 1 in-house on what people actually change in a draft, after separating mishearing from a change of mind. The evaluation is also candid: lowest word error rate on random real dictation, second on the hardest challenge set, with Wispr noting the winner there is a frontier model that cannot run in real time. Two limits matter in practice: meeting notes are Mac and Windows only, and the free tier caps words by the week. The number to judge it by is not accuracy but how often you have to edit.
Version History
- First on the Voice Arena dictation benchmark (2026-11-15): The independent Voice Arena Dictation Bench ranked Wispr Flow first, with 79.4 percent of drafts ready to send unedited and 0.40 percent of characters needing changes against 0.93 percent for the next best; when wrong, no draft needed more than 19 characters fixed. Wispr said Canto 2 is launching soon and will be judged the same way, by how often people have to fix what Flow writes.
- Canto 1 and the in-house route (2026-09-17): The Wispr Advanced Interfaces Lab released Canto 1, the company's first in-house speech model, led by Ariya Rastrow, a founding scientist behind Alexa, as chief science officer. Starting from a model pretrained on millions of hours of speech and text, it was trained in two stages of supervised fine-tuning and GRPO reinforcement learning, with the signal drawn from real user edits and mishearings separated from self-corrections. It posted the lowest word error rate on a 10 hour evaluation set drawn from more than 2,300 speakers, and placed second on a challenge set of nearby speech, traffic, whispers and short dictations, the lowest among real-time transcription models.