Does Wispr Flow run locally?

The name does most of the work. Whisper is OpenAI's open-source speech-to-text model, released in 2022, the one anyone can download and run locally. "Wispr Flow" lands close enough that people reasonably assume it's running Whisper, or that the two are related. They aren't.

The other source of confusion is feel. Wispr Flow is fast, and when transcription comes back that quickly it's tempting to assume the model must be running on the laptop. It doesn't have to be. A fast cloud call to a GPU farm is also quick, just not local. Wispr Flow is cloud-based, the audio uploads to their servers, and the app stops working without an internet connection.

Every time you dictate into a cloud tool, the audio leaves your machine. That has a few consequences worth being honest about:

• Privacy. Confidential client briefings, medical notes, anything you'd rather not have a copy of sitting on someone else's servers, gets a copy sitting on someone else's servers.
• Latency. It depends on your connection. On a train, on hotel wifi, on a flight, it slows down or stops entirely.
• Compliance. GDPR classifies voice as biometric data when it can identify a speaker. HIPAA fines for routing medical audio through unauthorised cloud services run up to $2.07M per violation.
• Pricing model. Subscriptions, because someone has to keep paying for the GPUs.

If "I want my dictation to actually be local" is what brought you here, Wispr Flow isn't what you're looking for.

You can run Whisper itself locally. Plenty of Mac apps do, including Whisper Notes, MacWhisper and SuperWhisper. The Large V3 model is around 3 GB on disk, runs through CoreML or whisper.cpp, and works offline.

Two problems show up once you do.

Speed. NVIDIA's published benchmarks put Whisper Large V3 at a real-time factor of 146x: 146 seconds of audio transcribed per second of compute on their reference hardware. Quick on paper, but for short push-to-talk bursts the model load and warm-up time dominates.

Hallucinations. This is the bigger issue. A peer-reviewed study presented at FAccT 2024 by Cornell researchers found that around 1% of Whisper transcriptions contain entirely fabricated phrases, and 38% of those fabrications include violent or otherwise harmful content. The same problem surfaced in hospital transcription pipelines and was covered in Wired. The fabrications appear during silences and between speakers, where Whisper's encoder-decoder architecture keeps trying to produce language even when there's nothing to transcribe.

So "just run Whisper locally" solves privacy but inherits Whisper's accuracy and reliability problems. The configuration story for SuperWhisper specifically (local Whisper, optional cloud models, BYOK to OpenAI or Anthropic) is in Parakeety vs SuperWhisper.

Parakeety, local Mac dictation that actually runs on-device, doesn't use Whisper. It uses NVIDIA's Parakeet TDT 0.6B v3, a different open model from a different lineage. It runs on the Apple Neural Engine, the dedicated machine-learning silicon on every Apple Silicon Mac.

The numbers, from NVIDIA's published benchmarks:

• Word error rate: Parakeet TDT v3 at 6.32%, Whisper Large V3 at 7.44%. A smaller, faster model that's also more accurate.
• Real-time factor: Parakeet TDT v3 at 3,333x, Whisper Large V3 at 146x. Roughly 23x faster on the same hardware.
• Model size: Parakeet TDT v3 is 0.6B parameters, Whisper Large V3 is 1.5B. About 2.5x smaller.

The third-party signal here is worth flagging. Whisper Notes, a Mac dictation app named after Whisper, switched its default engine to Parakeet TDT v3 in 2025. When the app named after Whisper has reasons to use a different model, that's a meaningful signal.

The hallucination story is structural rather than a quirk of training data. Parakeet uses a transducer architecture instead of encoder-decoder. Transducers can emit "no output" frames natively, so during silences they produce silence in the transcript rather than improvising sentences. The class of failure that makes Whisper unsuitable for unattended transcription doesn't apply.

• Wispr Flow: cloud, fast, requires internet, subscription, audio leaves your machine.
• Local Whisper: on-device, slower in practice, prone to hallucinations, free if you wire it up yourself.
• Parakeety (Parakeet TDT v3): on-device, faster than cloud Whisper, more accurate than Whisper Large V3, no hallucination class, $30 once.

Cloud transcription APIs sit at $0.21 to $1.44 per hour of audio depending on the provider. Parakeety is $30 once and includes every future update. If you dictate even an hour a week, the maths gets uncomfortable for the cloud option fairly quickly. For the broader feature comparison against Wispr Flow specifically (pricing, languages, the HIPAA posture) see Parakeety vs Wispr Flow.

Two honest caveats:

• Languages. Parakeet TDT v3 covers 25 European languages. If you dictate in Mandarin, Japanese, Arabic or anything outside that set, neither this model nor this app is the right tool yet.
• Specialist vocabulary. Any speech model can stumble on dense medical or legal jargon. Parakeet does well on this in independent tests, but if your dictation is mostly specialist terminology, expect to do some manual correction whichever engine you pick.

Parakeety is a Mac menu-bar app. Hold the section key, talk, release; your words paste at the cursor in whichever app you were typing into. Audio never leaves the machine. There's a free 7-day trial with no card required. After that it's $30 once.

Try Parakeety free →