Is ElevenLabs speech to text better than Whisper?

On the best independent test we found, ElevenLabs Scribe v2 is more accurate than Whisper, but that is not the same as a verdict for your audio. Scribe v2 led the Artificial Analysis AA-WER v2.0 benchmark at 2.3 percent ¹, and the article's key results do not include Whisper. We found no published test that runs both models on the same recordings. Whisper is also free and runs on your own computer, while Scribe v2 is a paid cloud service at $0.22 per hour ². So the answer depends on whether accuracy or cost, privacy and offline use matter more to you. For the wider comparison, see the guide to speech-to-text APIs.

Scribe v2 ranks first on the independent benchmark, but the test is not head to head

The Artificial Analysis benchmark is the most recent independent comparison we could find that includes Scribe v2. Its key results report that Scribe v2 leads at 2.3 percent, followed by Google's Gemini 3 Pro at 2.9 percent, Mistral's Voxtral Small at 3.0 percent, Google's Gemini 2.5 Pro at 3.1 percent and Gemini 3 Flash at 3.2 percent ³. We did not find Whisper in those key results.

The result has limits. Scribe v2 leads on two of the three datasets in the benchmark, and Google's Gemini 3 Pro leads on the parliamentary VoxPopuli dataset ⁴. A single benchmark with its own datasets and text rules tells you how the systems rank on that mix of audio, not how they will do on yours. The guide to speech recognition accuracy explains how these scores are built.

The open leaderboard puts Whisper at 7.44 percent, on a different test

The Open ASR Leaderboard is the other public comparison, and it does include Whisper. Its English short-form track lists OpenAI Whisper Large v3 at an average word error rate of 7.44 percent ⁵. That number cannot be placed next to the AA-WER score of 2.3 percent. The two tests use different audio, different normalization and different averaging, so both sets of figures stand on their own. Do not read the gap between them as a measurement of how much better Scribe v2 is.

Whisper is free to run and its code is MIT licensed

Whisper's code and model weights are released under the MIT License ⁶. You can download it, run it on your own hardware and pay nothing per hour of audio. The cost moves to the machine. The large model needs about 10 GB of video memory ⁷, so a computer with less memory has to use a smaller model.

Free and local is not the same as ready to use. You install the software, choose a model size, and handle the audio pipeline. The guide to running Whisper locally covers the setup, and the question on whether Whisper runs offline covers what needs a network connection.

The vendor's claims favor Scribe v2, and you should test them

ElevenLabs says that Scribe outperforms Gemini 2.0 Flash, Whisper Large V3 and Deepgram Nova-3 on the FLEURS and Common Voice benchmarks across 99 languages ⁸. That is a vendor's comparison on test sets it chose. ElevenLabs also states that Scribe v2 has over 98 percent transcription accuracy, but its pricing page gives no test set or method for that figure ⁹.

Treat both as reasons to test, not as results. Run each model on about five minutes of your own audio, with the names and terms you use, and calculate the word error rate by the same rules for both. The word error rate guide shows the calculation.

Whisper can invent text, so check for it

Accuracy is not only a count of wrong words. A 2024 study found that about 1 percent of Whisper transcriptions contained entire hallucinated phrases or sentences that were not in the audio ¹⁰. Invented text is harder to spot than a misheard word, because it reads as fluent speech. If you use Whisper for records, search its output for sentences that do not match the recording, and read the question on why Whisper hallucinates for the known causes.

Scribe v2 is a different system and has its own error profile, which the benchmark above does not show. Check both for invented text in your test.

It depends on cost, privacy, speed and how you will use the text

• Choose Scribe v2 when you want the top benchmark score without running a model yourself, or when you need speaker labels in the same service. ElevenLabs says its Scribe API returns speaker diarization with word-level timestamps ¹¹. Its rate is $0.22 per hour before add-ons ².
• Choose Whisper when audio must stay on your machine, when the volume is high enough that an API bill matters, or when you have the hardware and the staff to run a model.
• Choose neither for a single recording you need to read today. A free web tool such as our browser-based file transcription may be enough.

If you want dictation that runs on the device without a cloud call, Yap, which we make, is an on-device dictation app for Mac and Windows. You can download it and test it on the same recordings.

What to do

1. Pick five minutes of audio that includes your names, terms and background noise.
2. Transcribe it with Scribe v2 through the ElevenLabs API, and with Whisper on your own machine, using the same language setting.
3. Count word errors and invented sentences for each, using the same normalization.
4. Choose the cheaper system that stays within your error limit and your data rules.

Related answers: how to use the ElevenLabs speech-to-text API, whether Whisper has an API, whether Whisper AI is free and which Whisper model is best. The overview of speech recognition explains the technology both systems use.