Fact · Speech recognition
6.66percent
On the Hugging Face Open ASR Leaderboard (its version of October 17, 2026, read on October 20, 2026), openai/whisper-large-v3 had an average word error rate of 6.66 percent across the leaderboard's 9 public English test sets.
Read by us from Hugging Face Hub (dataset hf-audio/open-asr-leaderboard-results) on October 20, 2026
Where this number was read
This number was read by our collector, word for word, from an official public source, and the reading is kept with every earlier one. Open the exact request below to see the figure as it stands today.
- Field
- mean of AMI-Cleaned WER, Earnings22-Cleaned-AA-chunked WER, Gigaspeech-Cleaned WER, LS Clean WER, LS Other WER, SPGISpeech WER, URGENT2024 WER, Voice Arena Monsoon WER, Voxpopuli-AA-Cleaned WER
- Read
- October 20, 2026, and again every week
- What it measures
- The Open ASR Leaderboard, run by Hugging Face, scores open models and cloud APIs on the same public English recordings with one evaluation script and the same text normalization, so its figures compare models with each other rather than predict accuracy in one person's dictation.
- How it is measured
- Once a week we read the Open ASR Leaderboard's own configuration (its latest version, the exact revision of the English results file it shows, and its default test sets), fetch that results file from Hugging Face at that revision, and record for every model its word error rate on each public default test set and their average (the mean, rounded to two decimals, as the leaderboard computes it). A model without a score on every one of those sets is left out, as the leaderboard leaves it unranked. Speed (RTFx), license and size are the leaderboard's columns as published.
- What it does not show
- The leaderboard's own Average WER also includes two averages over private recordings (scripted and conversational) that are not published, so its ranking can differ slightly from this public-set average. The test sets are read audiobooks (LibriSpeech), meetings (AMI), earnings calls (Earnings-22, SPGISpeech), podcasts and videos (GigaSpeech), European Parliament speeches (VoxPopuli), degraded and noisy speech (URGENT 2024) and Indian English conversations (Voice Arena), not dictation; text is normalized before scoring, so capitals and punctuation are not counted. Speed is measured on the leaderboard's own hardware, and cloud APIs have no speed figure.
In the data
A row of Open ASR Leaderboard: Speech Recognition Accuracy, Tracked, as openai/whisper-large-v3: 6.66 percent.
- Model
- openai/whisper-large-v3
- Maker
- openai
- Access
- Open weights
- Product
- Whisper
- License
- apache-2.0
- Size (billion parameters)
- 2
- Speed (RTFx)
- 470.73
- LibriSpeech clean
- 1.56
- LibriSpeech other
- 3.52
- AMI meetings
- 13.63
- Earnings-22 calls
- 7.99
- GigaSpeech
- 8.35
- SPGISpeech
- 2.71
- VoxPopuli
- 4.53
- Noisy speech (URGENT 2024)
- 13.73
- Indian English conversations
- 3.95
- Leaderboard version
- 2026-10-17
Used to calculate
Cited on
- Speech Recognition Benchmarks and LeaderboardsGuide
- Are open-source speech recognition models as accurate as paid APIs?Question
- Is on-device speech recognition as accurate as cloud?Question
- Whisper or Parakeet: which is more accurate?Question
- Open ASR Leaderboard: Speech Recognition Accuracy, TrackedDataset