Benchmarks

Which Whisper model fits your PC?

The right Whisper model is the one that handles your recordings without slowing down the rest of your PC. Start with the local model installed in WhisperARC. Then compare a short sample before downloading a heavier option.

By WhisperARCUpdated

Silver processors with a small engraved audio waveform against a smooth black background

Use published comparisons as a starting point

OpenAI's reference table compares Whisper model families on English speech using an NVIDIA A100. It puts Turbo ahead of Small in that setup despite Turbo having more parameters. Model size alone does not determine speed.

WhisperARC runs local models through whisper.cpp on Windows. Its quantized files and your hardware differ from the reference test. Use the chart to understand the comparison, then check the result on your own PC.

At a glance

A smaller model is not always faster

Approximate relative speeds reported by OpenAI for English speech on an NVIDIA A100. These are reference-model results, not WhisperARC measurements on Windows.

Large · reference1×
Large is the 1× baseline in OpenAI's table. This is a relative comparison, not seconds of processing per minute of audio.
Small4×
Small: approximately 4× the reference speed in this specific test. The model family has 244 million parameters.
Turbo8×
Turbo: approximately 8× the reference speed, despite having more parameters than Small. Architecture matters alongside size.

Read the conditions with the number: a different runtime, language or machine can change the ordering.

Hover or focus a highlighted element for the detail.

Start with the workload you actually have

For short dictations you notice the pause after releasing the shortcut. For an interview you care more about how long the whole recording takes. Test the kind of work you plan to do.

Open Model > Local > Transcription to see the available files. Base q8 and Small q8 have the smallest downloads in the current catalogue. The Large v3 Turbo variants offer another place to compare quality and processing time. The hardware guide lists their storage and memory recommendations.

Compare the same recording

Choose a clear sample that includes the names and technical terms you use. Run it with your current model, then repeat it with one alternative. Keep the recording and processing settings the same.

Note when the text becomes usable. Then read it against the recording and count the errors that need attention. A quicker first draft can lose its advantage if you spend longer checking a surname or repairing a sentence.

  • Include a short dictation and a longer recording if you use both.
  • Check the first run and a repeat run with the model already loaded.
  • Keep other heavy applications open if that reflects your working day.

Check compute and memory before moving up

Under Model > Local you can choose CPU processing or a compatible Vulkan GPU. The model still needs enough working memory. Its download size is not a measurement of the RAM or graphics memory it will use.

If the PC becomes sluggish, compare a lighter model or adjust the compute settings. If the same important words keep failing, try another model on that sample. Keep the option that gives you reliable text with a manageable delay.

Sources & further reading

Keep exploring.

All articles