OpenAI

Whisper on your own PC: hardware for transcription

Published Oct 27, 2026 Updated Oct 27, 2026

We earn from qualifying purchases via affiliate links. Disclosure.

Verdict

OpenAI's Whisper README lists about 10 GB of VRAM for large and about 6 GB for turbo, the default. Any 12 GB GPU covers both. Smaller models (tiny to medium) need about 1–5 GB.

Prices: Newegg checked Oct 27, 2026, about 12:35 PM CT; Minisforum checked Oct 27, 2026, 12:40 PM CT. New condition only.

Model memory at a glance

Model	Memory needed	Format	Context	Smallest tier (our estimate)	Source
Whisper large / turboOpenAI · Speech	~10 GB / ~6 GB VRAM	as published	—	8–12 GB GPU8–12 GB GPU	OpenAI Whisper README (model sizes and VRAM)

Good, Better, Best for this model (8–12 GB tier)

Three new-retail picks, chosen from cited specs and prices. We do not test hardware ourselves. Prices show when we checked them.

Good

Cheapest 12 GB card we priced

ASRock Challenger Intel Arc B580 12GB

$329.99 Newegg · checked Oct 27, 2026, about 12:35 PM CT

Holds 8B–14B chat models at Q4 and SDXL (per the sources). Needs a desktop to go in; check your runtime supports Intel Arc.

Sold by Newegg, seen on the product page

Check price at Newegg

Better

Step up to 16 GB

ASUS Dual GeForce RTX 5060 Ti 16GB OC

Price shown at Amazon

4 GB more VRAM than the Good pick, which opens up gpt-oss 20B (14 GB download). Needs a desktop to go in.

Sold at Amazon.

Check price at Amazon

Best

ABS Cyclone series mesh mini-tower case with RGB fans — illustrated product card

Complete PC, nothing to build

ABS Cyclone Aqua gaming PC (Core i7-14700F, RTX 5060 Ti 16GB, 16GB DDR5, 1TB SSD)

$1,599.99 Newegg · checked Oct 27, 2026, about 12:35 PM CT

One box with the 16 GB card already inside. Ships with 16 GB system RAM; plan an upgrade to 32 GB.

Sold by Newegg, seen on the product page. Ships with 16 GB system RAM; plan an upgrade to 32 GB.

Check price at Newegg

The details

Whisper model sizes

Model	Parameters	Required VRAM	Relative speed
tiny	39 M	~1 GB	~10x
base	74 M	~1 GB	~7x
small	244 M	~2 GB	~4x
medium	769 M	~5 GB	~2x
large	1550 M	~10 GB	1x
turbo	809 M	~6 GB	~8x

All figures are from the Whisper README. OpenAI measured the speeds relative to large by transcribing English speech on an A100, and says real-world speed varies a lot. The README also says turbo is not trained for translation, so use a multilingual model to translate.

Which GPU

A 12 GB card runs every size. If you want Whisper and a chat model on the same GPU at once, add the two memory figures together and pick a tier from the model finder.

Install and run Whisper

Steps and commands come from the official docs listed at the end of this section, checked Oct 27, 2026, 2:30 PM CT. We don't add flags the docs don't show. Commands can change, so check the linked page if one fails.

1. Before you start

whisper.cpp builds with CMake. For NVIDIA, install CUDA first. For Vulkan, your graphics driver needs to support the Vulkan API (Intel Arc, AMD and NVIDIA drivers do). For AMD ROCm, install ROCm. On Apple silicon, whisper.cpp runs fully on the GPU through Metal.

2. Get whisper.cpp and a model

git clone https://github.com/ggml-org/whisper.cpp.git
cd whisper.cpp
sh ./models/download-ggml-model.sh large-v3-turbo

3. Build for your GPU

Hardware	Build command (whisper.cpp README)
Apple silicon / CPU	`cmake -B build` then `cmake --build build -j --config Release`
NVIDIA	`cmake -B build -DGGML_CUDA=1` then `cmake --build build -j --config Release`
Intel Arc, AMD or NVIDIA via Vulkan	`cmake -B build -DGGML_VULKAN=1` then `cmake --build build -j --config Release`
AMD via ROCm	`cmake -B build -DGGML_HIP=1 -DAMDGPU_TARGETS="gfx1201"` then `cmake --build build -j --config Release`. Replace `gfx1201` with your GPU's architecture.

4. Transcribe

whisper-cli takes 16-bit WAV, so convert your audio first:

ffmpeg -i input.mp3 -ar 16000 -ac 1 -c:a pcm_s16le output.wav
./build/bin/whisper-cli -m models/ggml-large-v3-turbo.bin -f output.wav

5. Check the GPU and fix problems

Watch nvidia-smi (NVIDIA) or rocm-smi (AMD) while it runs.
Out of memory: download a smaller model (the README lists small, medium, large-v3-turbo and more) and pass it with -m.
Run ./build/bin/whisper-cli -h for all options.