Whisper: the incumbent everywhere
OpenAI released Whisper in 2022 and accidentally created the modern dictation industry. A single open model, trained on 680,000 hours of web audio, recognising 99 languages at quality that beat most commercial systems — and free for anyone to ship. The whisper.cpp port made it run on ordinary laptops, and nearly every local dictation app in existence is a wrapper around it: Handy, Superwhisper, MacWhisper, VoiceInk, SpeechPulse, OpenWhispr and more.
Whisper's strengths are breadth and robustness: it handles accents, noise and code-switching training data threw at it, and its language coverage remains unmatched in the open world. Its weaknesses are architectural age. It hallucinates text during silence or music (producing confident transcripts of things nobody said), it repeats itself on looped audio, its 30-second window forces chunk-and-stitch pipelines, and it is slow relative to 2026 alternatives — large-v3 quality costs real seconds per chunk on modest hardware.
Parakeet V3: the local challenger
NVIDIA's Parakeet TDT 0.6B V3 arrived in 2026 and redrew the local tradeoff. It covers only 25 languages — European ones, plus Russian and Ukrainian — but within that list it beats Whisper large-v3 on accuracy benchmarks while running several times faster on the same hardware and fitting in much less memory. For an English, Spanish, French or German speaker, a Parakeet-based app gives large-model quality at small-model speed.
The catch is exactly the list. Dictate in Korean, Arabic or Hindi and Parakeet simply isn't an option; apps that ship it (VoiceInk is the prominent example) keep Whisper builds as the fallback for the long tail. Parakeet also punctuates better than raw Whisper out of the box, which narrows one of the cloud apps' traditional advantages.
The proprietary cloud models
At the top of the market, Wispr Flow, Aqua Voice and Willow Voice run their own models — trained or heavily fine-tuned in-house, served from their data centres, and updated continuously without app releases. This is the biggest structural shift of the past two years: the leaders no longer compete on whose Whisper wrapper is nicer, they compete on proprietary recognition that the open ecosystem doesn't have.
What proprietary buys: better accuracy on messy real-world audio, proper streaming with self-correcting partials, automatic language detection, and per-language formatting tuned by teams who can measure their own error rates. What it costs: your audio leaves your machine, quality depends on their uptime, and the model can change underneath you — for better or worse — with no changelog.
How to tell which engine an app runs
Vendors are oddly shy about this, but the tells are reliable. Works fully offline with a model download of 1–3 GB: Whisper or Parakeet. Requires an account and a network connection at all times: proprietary cloud. Offers a model-size picker in settings: Whisper, certainly. Lists exactly 25 languages: Parakeet. Claims 'over 100 languages' with automatic detection and no offline mode: proprietary built on Whisper-scale training.
Our per-app reviews name the engine explicitly, because it predicts so much: offline capability, language ceiling, latency character, and the shape of the privacy story.
References
- OpenAI — Whisper — the model that created the category
- NVIDIA — Parakeet TDT 0.6B V3 model card — 25 languages, benchmarked against Whisper large-v3
Frequently asked questions
Is Parakeet always better than Whisper?
Only inside its 25-language list, where it's faster and at least as accurate. For anything outside that list — Asian languages, most of the world — Whisper remains the open standard.
Will open models catch up to the proprietary ones?
They have so far, with a lag of roughly one to two years. The pattern since 2022: a proprietary lead appears, an open release closes most of it, repeat. The cleanup-layer models are where the gap is currently widest.