TL;DR: For push-to-talk dictation, any NVIDIA card with about 2 GB of free VRAM handles the default model in real time. A GTX 1060 is enough. What kills performance isn’t a weak GPU, it’s VRAM pressure from everything else you’re running, or accidentally landing on the CPU path without noticing.
Dictation is a small workload
People see “AI transcription on GPU” and picture the hardware requirements of running an LLM. It’s not that. A dictation clip is five to thirty seconds of audio, and Whisper models are small compared to language models. The default model I ship in Invoke, distil-large-v3, is under a billion parameters. At float16 that’s roughly 1.5 GB of weights.
So the honest requirements table is short. These are the numbers I see in practice, rounded, with the model loaded and transcribing:
| Model | VRAM in use | Good for |
|---|---|---|
| tiny / base | under 1 GB | Old cards, laptops |
| small | ~1 to 1.5 GB | Budget GPUs, decent accuracy |
| distil-large-v3 | ~2 GB | The sweet spot for dictation |
| large-v3 | ~4 GB | Long recordings where accuracy is everything |
Your exact numbers will wander a few hundred megabytes either way depending on compute type and runtime. The shape holds: this fits on almost anything.
What speed looks like per tier
On my desktop (an RTX 5090, which is absurd overkill for this) a short utterance comes back in well under half a second. But the interesting number is how little card you need before dictation stops feeling instant.
A mid-range card like a 3060 or 4060 transcribes a short clip in a few tenths of a second. An eight-year-old GTX 1060 still lands around a second, which in a push-to-talk flow reads as “immediate.” The release-to-text gap you can feel starts on the CPU path, where the same clip takes several seconds and you sit there watching the cursor.
That’s the actual dividing line in local dictation: not which GPU, but GPU versus no GPU.
VRAM headroom is the real spec
Here’s the failure mode nobody warns you about. The model needs its 2 GB while everything else is also using VRAM. Chrome takes some. Your IDE takes some. A game takes most of it.
If the transcription model can’t get memory, you don’t get a clean error, you get slowness or a fallback. So when dictation that was instant yesterday feels sluggish today, check what else is sitting on the card before blaming the software. On an 8 GB card this basically never bites. On a 4 GB card running a game, it can.
The other thing worth knowing: the model loads into VRAM once, at startup, and that load takes a couple of seconds. Good dictation software keeps it resident so every press of the hotkey hits a warm model. If a tool reloads the model per utterance, no GPU on earth will make it feel fast.
Invoke is $49 once. Free 7-day trial, all features, no credit card.
No NVIDIA card?
The local-GPU story is CUDA-shaped, and the alternatives are mixed:
Apple Silicon is fine. Whisper runs well on the M-series GPU through Metal and MLX. It’s the one non-NVIDIA platform where local dictation is genuinely first-class.
AMD and Intel Arc are rough for this specific workload. The fast Whisper runtimes are built on CUDA, and the workarounds (Vulkan builds of whisper.cpp, ROCm experiments) exist but you’re off the paved road. If you have a Radeon card, expect to run the CPU path.
CPU only works, with the right expectations. Use a small model with int8 quantization and clips come back in a couple of seconds. Fine for occasional use, wearing for all-day dictation. I wrote up the CPU-versus-GPU setup details in how to run Whisper locally on Windows.
How to check which path you’re on
The sneakiest problem in local dictation is having a perfectly good GPU and not using it. Broken CUDA library installs fail quietly and fall back to CPU, and everything still works, just four times slower. If dictation takes seconds on a machine with an NVIDIA card, that’s the first thing to check. In Invoke, dictate health prints exactly which device and model you’re on, because I got tired of guessing.
The deeper reasons a local GPU beats cloud transcription for developers (latency, privacy, cost) are a separate argument, and I made it in local voice to text for coding.
The practical answer
If you bought a gaming or workstation NVIDIA card any time since about 2016, you already own dictation hardware. The model is small, the clips are short, and the GPU spends 99% of its day idle waiting for you to talk.
Invoke sets this up without the CUDA scavenger hunt: the libraries ship in the installer, the model loads warm at startup, and you find out in the first minute whether your card is up to it, because the trial is the full app. $49 once if it sticks. Free 7 days here.