Blog

Wispr Flow's Canto: what's new and how blabe compares

What Wispr has published about Canto, its speech model for dictation, what nobody has measured yet, and how blabe and the speech models it uses compare.

On September 17, 2026, Wispr introduced Canto, which it calls "our latest speech model for real-time dictation". We make blabe, a lightweight and extremely fast dictation app for Mac and Windows, so people ask us how the two compare. Below is what Wispr has published, what nobody has measured yet, and where blabe is different. Every claim about Wispr links the page it came from, read on November 2, 2026.

What Canto is

  • Wispr's own model. It comes from the Wispr Advanced Interfaces Lab and is the first model in a broader research and development program, according to Wispr's Canto page.
  • Tested on English dictation. Wispr built its main test set from 10 hours of English Wispr Flow dictations by more than 2,300 speakers who opted in to data sharing. On that set, Wispr reports that Canto had the lowest word error rate of the models it compared, from Google, OpenAI, AssemblyAI and Deepgram.
  • Hard audio. On a separate 3-hour set of difficult recordings (other voices, music, traffic, wind, quiet or whispered speech, distant microphones and one- or two-word dictations), Wispr reports that Canto came second to Gemini 3.1 Pro and first among the real-time models it tested.
  • Public benchmarks. On the English parts of three public test sets, Wispr says Canto tied for the lowest error rate on LibriSpeech and "remained competitive on FLEURS and Common Voice, although it did not lead either evaluation."
  • Your vocabulary. While you dictate, Canto is given the words in your Wispr Flow dictionary to help it with names and rare terms, Wispr explains.

What isn't known

  • Canto hasn't been tested side by side with the models blabe uses. Wispr's comparison covers models from Google, OpenAI, AssemblyAI and Deepgram. It doesn't include xAI's Grok Voice Transcribe 2.0 or Microsoft's MAI-Transcribe 2. As of November 2, 2026, we know of no independent public comparison of Canto with either.
  • The published numbers don't line up. Wispr's FLEURS result is on the English part of the test. Microsoft says MAI-Transcribe 2 ranks #1 on FLEURS at 5.2% WER across 60 languages. One is a single language and the other an average across 60, so they can't be compared directly.
  • Your own speech is the real test. Neither company's numbers tell you how a model handles your voice, your microphone and your project's names. Wispr Flow lists a $0 plan and blabe has a 2-day trial, so test them on the words you actually say.

How blabe compares

These are blabe's own figures; see Speed for the method. Speed and idle memory were measured on October 14, 2026, on blabe 1.0.21 for Mac and 1.0.4 for Windows, with MAI-Transcribe 2. That was before live transcription with xAI came in blabe 1.0.24 for Mac and 1.0.8 for Windows. The download sizes are for blabe 1.0.27 on Mac, measured October 24, and 1.0.14 on Windows, measured October 24.

  • About 1 s from speech to text. Across 43 real dictations in the Mac app, text was pasted 1.0 s after the end of recording at the median, with AI cleanup on.
  • Small and light. A 4.6 MB download on Mac and 12.5 MB on Windows. One process, ~75 MB idle on Mac, ~22 MB on Windows.
  • The speech models. Settings → Dictation → Transcription has two options. With xAI (live), the default, most dictation in English or with no pinned language streams to xAI's Grok Voice Transcribe 2.0 while you speak, and the rest goes to Microsoft's MAI-Transcribe 2 on Azure. With Azure, your audio goes only to Microsoft Azure. Both run through blabe's own transcription service and are included in the subscription, with no provider account or API key to set up. Our privacy page has the details.
  • Your words as transcribed, with AI cleanup off. AI cleanup is Aggressive, Standard or Off, and Aggressive is the default for new installs. With it off, your words are pasted as transcribed and no language model rewrites them. Wispr Flow removes fillers, adds punctuation and formats lists, which suits you if you want polished text rather than a transcript.
  • Where it runs. blabe runs on Mac (macOS 14 or later) and Windows 11 on x64 PCs. Wispr Flow runs on Mac, Windows, iPhone and Android.
  • Price. blabe is $7.99 a month or $69 a year for up to 3 devices, with every feature included, after a 2-day trial with no sign-up. Wispr Flow's current plans are in the side-by-side table.
  • Internet. Both need a connection to transcribe. Wispr says "an internet connection is required for transcription", and so does blabe.

Which one to pick

Pick Wispr Flow if you want your speech turned into polished writing, if you also dictate on your phone, or if Wispr's English results on difficult recordings match how you work.

Pick blabe if you want your words as transcribed in any app with AI cleanup off, a lightweight app that stays out of the way, and one plan for your Macs and Windows PCs. Our comparison of blabe and Wispr Flow has the full side-by-side table, with a source for every Wispr cell.

blabe is not affiliated with or endorsed by Wispr. Wispr Flow and Canto are names used by their owner. You can try blabe free for 2 days with no sign-up, card or email; after that it's $7.99 a month or $69 a year (pricing).