Updated November 2026
Whisper benchmarks on Windows.
Whisper speed on CPU and GPU, from public tests.
What to pick
With a graphics card
Large v3 Turbo. It is close to Large v3 in accuracy and 2 to 3 times faster.
Without a graphics card
Base or Small for dictation. Large models can take more than 10 times the audio length on a CPU.
Short on disk space
A Q8 or Q5 version. The file is much smaller and accuracy barely changes.
CPU vs GPU
11-second clip · Core i9-13980HX, RTX 4090 Laptop.
| Model | CPU | GPU · Vulkan | GPU · CUDA |
|---|---|---|---|
| Tiny | 1.98 s | 0.31 s | 0.32 s |
| Base | 4.54 s | 0.42 s | 0.41 s |
| Small | 20.4 s | 0.82 s | 0.84 s |
| Medium | 73.6 s | 2.02 s | 1.91 s |
| Large v3 | 133.5 s | 3.73 s | 3.90 s |
Source: StarWhisper, July 2026
Turbo vs Large v3
Speed versus audio length. Higher is faster.
| Test | Large v3 | Large v3 Turbo |
|---|---|---|
| RTX 4070 Ti · whisper.cpp | 15× real time | 40× real time |
| NVIDIA A100 · OpenAI | 1× (reference) | About 8× |
Source: x2q, June 2026 · OpenAI Whisper
Integrated graphics
Large v3 Turbo Q8 · Ryzen 7 PRO 4750U laptop.
| Clip | CPU | GPU · Vulkan |
|---|---|---|
| 11 seconds | 9.85 s | 4.40 s |
| 35 seconds | 20.36 s | 9.59 s |
Source: transcribe.cpp, September 2026
Smaller files, same accuracy
Large v3 Turbo on LibriSpeech. Lower is better.
| Version | Word error rate |
|---|---|
| Full (F16) | 2.01% |
| Q8 | 2.01% |
| Q6 | 2.01% |
| Q5_K_M | 2.03% |
| Q4 | 2.04% |
Source: transcribe.cpp, September 2026
Whisper benchmarks FAQ
Do I need a graphics card to run Whisper?
No. Base and Small run on a recent processor. A graphics card makes large models much faster: 36 times faster in one test with Large v3.
Is Vulkan as fast as CUDA for Whisper?
On NVIDIA cards, yes. In the same test, Vulkan and CUDA stayed within a few percent of each other. Vulkan also works on AMD and Intel graphics.
Which Whisper model is the fastest?
Tiny and Base are the fastest. Among the accurate models, Large v3 Turbo is much faster than Large v3.
Do quantized Whisper models lose accuracy?
Very little. On clean English speech, the error rate went from 2.01% for the full version to 2.04% for the smallest one.