Is Whisper more accurate than Google Speech-to-Text?

Not on every test. In one 2024 study of English lecture recordings, Whisper large-v2 made far fewer word errors than Google's service. A 2024 classroom study found Google more robust to noise, and speech models change often, so a result from one test, one audio type and one year does not settle the question for your work.

One 2024 test of English lectures gave Whisper the lower error rate

Kuhn and colleagues measured eleven speech recognition services on recordings of university lectures ¹. On the English recordings, Google's average word error rate (WER) was 20.1 percent ². Whisper large-v2 averaged 2.9 percent on the same recordings ³. Google also had the highest mean error rate of the services tested, and the widest spread between recordings ¹.

Those are large gaps, and they come from a single set of lecture recordings, so the result describes that kind of audio. The same paper's conclusion is that accuracy varies between services and between individual recordings, which means one average hides a lot of spread.

For the clean-audio number that OpenAI itself published, Whisper large-v2 had a WER of 2.7 percent on LibriSpeech Clean, a benchmark of read audiobook speech ⁴. A clean benchmark is useful for comparing models, but it is the easy case.

Other tests give Google the edge on noisy audio

A 2024 preprint studied children's classroom dialogue. The authors ran the recordings through Google Speech-to-Text, Rev.ai and Whisper, and they wrote that Google "achieved the most robust performance in the face of substantial audio noise" ⁵. The paper reports its error scores on its own scale, so compare them only with each other, and treat the study as a narrow data point about children's classroom speech.

Put the two studies together and the pattern is not a ranking. Whisper led on the lecture recordings, and Google led on the noisy classroom audio. Your own noise, speakers and vocabulary decide which result applies to you.

Vendor benchmarks are claims to test, not independent comparisons

ElevenLabs says its Scribe model "consistently outperforms leading models like Gemini 2.0 Flash, Whisper Large V3 and Deepgram Nova-3" on FLEURS and Common Voice across 99 languages ⁶. Gemini is a Google model, but Google's Speech-to-Text is not named on that list. The benchmark charts are images, and the page gives no test method in the text. Use vendor charts to decide which products to test, not as the answer.

Prices and data terms can decide the choice before accuracy does

Google Cloud's V1 list price for standard recognition with data logging is $0.016 per minute after the first 60 minutes each month ⁷. OpenAI's price list shows its hosted gpt-transcribe model at $0.0045 per minute ⁸. That is a newer OpenAI model, not the open Whisper weights, so check the model name before you budget.

Running Whisper yourself has no per-minute fee, but it costs hardware and setup time. Prices and terms change, so confirm them on the provider's current page on the day you decide. Our comparison of speech-to-text APIs lists more providers.

What to do: run both on your own audio

1. Record 10 to 20 minutes of the audio you will actually use, with the same microphone, room and speakers.
2. Write a reference transcript by hand, or correct a machine transcript while you listen.
3. Run each candidate on the same file with the same settings. For Google, use the model you would really buy, because older and newer models differ.
4. Compare the word error rate, and look at the errors that matter most, such as names, numbers and specialist terms.
5. Check the price and data terms for the monthly volume you expect before you commit.

The answer changes with the audio, the language and the model

If your audio is clean, English and read aloud, both services may perform well, and the cost decides. If it is noisy, multi-speaker or in a language with less training data, the gap can move either way, which is why a short test on your own recordings beats any published table.

For what Whisper is and where it struggles, read our guide to Whisper AI. For the score itself, see word error rate. If you need speaker labels as well as words, whether Whisper can identify different speakers covers that separate step.