Parakeet vs Whisper vs Nemotron: The Best Local Speech-to-Text in 2026

For English and major European languages, NVIDIA Parakeet TDT 0.6B v3 is the best local speech-to-text model in 2026: it beats Whisper large-v3 on measured accuracy at a quarter of the size, and runs dramatically faster on a plain CPU. If you want text that appears live while you speak, NVIDIA's new Nemotron streaming models are the first local models built for exactly that. And if you speak one of the many languages the NVIDIA models don't cover, OpenAI's Whisper — with 99 languages — is still the one to use.

The surprise of 2026 is that Whisper no longer holds the open-model accuracy crown it wore for three years. But “which model is best” has become a genuinely three-way question, because the models are good at different things: Parakeet optimizes accuracy-per-watt, Nemotron optimizes latency, and Whisper optimizes breadth. This article compares them with numbers from the models' own cards and the public leaderboard — no vibes, and no pretending the differences are bigger than they are.

Last updated July 18, 2026. We build OpenWhispr, an open-source dictation app that ships all three model families, so we benchmark and support each of them in production. That's also our bias — we don't sell any one of these models, and the honest trade-offs below are the same ones we explain to our own users.

• Parakeet TDT 0.6B v3 averages 6.34% word error rate across the eight Open ASR Leaderboard benchmarks, with a throughput of 3,332× real time. 600M parameters, 25 languages, CC-BY-4.0 license. Parakeet TDT v3 model card · Open ASR Leaderboard
• Whisper large-v3 averages roughly 7.4% on the same leaderboard and supports 99 languages with 1.55B parameters under an MIT license. OpenAI Whisper repository · Whisper large-v3 model card
• Nemotron Speech Streaming EN 0.6B averages 6.93% WER while streaming (1.12-second chunks), reaches 2.32% on LibriSpeech test-clean, and was trained on 530,000 hours of audio. Nemotron Streaming EN model card
• Nemotron 3.5 ASR covers 40 language-locales in one 600M streaming model with automatic language detection, released June 4, 2026 under the OpenMDW license. Nemotron 3.5 ASR model card
• All disk sizes and in-app behavior referenced below are verifiable in OpenWhispr's open-source model registry. OpenWhispr on GitHub · OpenWhispr models page

All three are free, open-weight models you can download and run offline. They differ in architecture, and the architecture explains almost every practical difference you'll notice.

Speech models are measured by word error rate (WER)— the percentage of words the model gets wrong. Lower is better. The standard reference is the Hugging Face Open ASR Leaderboard, which averages WER across eight diverse English datasets (meetings, earnings calls, TED talks, audiobooks, and more), so a single easy dataset can't flatter a model.

Read the chart honestly: every model here is production-grade, and the whole spread is about 1.5 points. In day-to-day dictation, a 6.3% and a 7.4% model feel similar on clean audio — the differences show up on hard audio (accents, noise, jargon). What makes the NVIDIA numbers notable is what you pay for them: Parakeet reaches its accuracy at 600M parameters versus Whisper's 1.55B, and Nemotron posts its 6.93% while streaming, without seeing the future audio at all.

On the leaderboard's hardware, Parakeet TDT v3 transcribes at 3,332× real time— an hour of audio in about a second. Whisper large-v3 posts well under a tenth of that. The reason is architectural, and it's worth understanding in plain terms: Whisper writes its transcript one token at a time, each step waiting on the last, while a transducer reads the audio once and emits text as it goes. No loop, no per-word round trips.

On your laptop the absolute numbers are smaller but the ratio holds: a dictated paragraph that takes Whisper large several seconds to process lands near-instantly with Parakeet — on CPU, with no GPU involved. Independent benchmarks also report the NVIDIA transducers using roughly 8–10× less energy per hour of audio than Whisper at comparable quality — relevant if you dictate all day on battery.

Whisper's answer to this is turbo(1.6GB), a distilled large-v3 that trades a little accuracy for roughly 8× Whisper-large speed. It's the right Whisper for dictation — but it narrows the gap rather than closing it. For the full Whisper size-by-size breakdown, see our Whisper model sizes guide.

The rule of thumb is simple. English or a major European language: any of the three, so pick on speed. Japanese, Korean, Vietnamese, Arabic, Hindi: Nemotron 3.5 now covers them with streaming — a real first for local models. Mandarin, Thai, Indonesian, Swahili, or heavy code-switching between languages: Whisper is still the only serious option, and that's unlikely to change soon.

Streaming is where Nemotron isn't just incrementally better — it's doing something the other two structurally can't. Whisper processes audio in 30-second windows and re-encodes everything it sees, so “live” Whisper is really the same batch model run over and over on overlapping audio — expensive, laggy, and jittery at the seams. Nemotron's cache-aware architecture was trained for the streaming case: it processes only the newest chunk of audio, carries its internal state forward, and emits partial text within a configurable latency budget of 80 milliseconds to 1.12 seconds.

The efficiency difference is dramatic at scale — NVIDIA reports 17× more concurrent streams than its previous-generation streaming model on the same GPU — but the part you actually feel on a laptop is simpler: text appears as you speak, from a model running on your own CPU. We wrote a full engineering breakdown of what it took to ship that in a desktop app in our streaming ASR deep dive.

If you don't care about live text — you dictate, stop, and want the final transcript — streaming buys you nothing, and Parakeet's batch models are the better pick. Streaming exists for live captions, voice agents, and the see-it-as-you-say-it dictation preview.

None of these models needs a GPU. The NVIDIA models ship as INT8-quantized ONNX files (631–680MB) that run comfortably on any recent CPU — Apple Silicon or x86 — via the sherpa-onnx runtime, a native binary with no Python and no PyTorch. Whisper runs everywhere through whisper.cpp; its small and turbo sizes are fine on CPU, while large-v3 (3GB, ~10GB RAM) genuinely benefits from acceleration.

If you do have a GPU, use it for Whisper: Metal is built in on Apple Silicon, and OpenWhispr offers one-click CUDA (NVIDIA) and Vulkan (AMD and Intel) runtimes with automatic CPU fallback if the GPU runtime fails. The NVIDIA transducers are fast enough on CPU that GPU acceleration is beside the point for dictation.

And if privacy is the reason you're going local in the first place, all three qualify equally: the audio is processed on your machine and never uploaded. We've written more about that trade-off in local vs cloud transcription.