benchmarks.site

Public leaderboard · Speech recognition (ASR)

Open ASR Leaderboard

Tests how accurately speech-to-text systems write down what people say in English recordings (meetings, earnings calls, audiobooks and more). Lower error means fewer wrong words.

Leaderboard

Top 3 right now

  1. 1stzoom/scribe_v2.1_pro4.87
  2. 2ndOtterAI/speech-v3.15.21
  3. 3rdyoob/yoob-asr-pro-v15.33

What it measures

Word error rate of open and proprietary speech recognition models across English, multilingual and long-form test sets, plus inverse real-time factor for speed.

How to read the numbers

Average word error rate across English datasets (%, lower is better) — lower is better.

Results

25 rows · lower is better (Average WER (%))CSV
1zoom/scribe_v2.1_pro4.87
2OtterAI/speech-v3.15.21
3yoob/yoob-asr-pro-v15.33
4microsoft/azure-speech-09-20265.5
5Musci-research/Musci-ASR-qwen-1019-hf5.7774.69
6p99lab/audrey-0.75.7
7elevenlabs/scribe_v25.93
8reson8/rumble-25.95
9modulate/multilingual5.96
10salesforce-sovereign-intelligence/SOV-Transcribe-15.96600.92
11FluidVoice-Labs/fluid-speech-pro6.063690.47
12sophea/asr-k16.11
13nvidia/NVIDIA-Nemotron-3.5-Transcribe-30B-A4B6.13472.62
14Qwen/Qwen3-ASR-1.7B-hf6.15835.62
15assemblyai/universal-3-5-pro6.21
16smallestai/pulse6.57
17ibm-granite/granite-speech-5.0-2b6.68548.43
18CohereLabs/cohere-transcribe-03-20266.72923.71
19HojoAI/Hojo-ASR-V16.7473.39
20TheStageAI/thewhisper-large-v3-turbo6.741183.92
21Edge0/ARK-ASR-0.6B6.75664.56
22openai/gpt-transcribe6.8
23sprag/symphony6.85
24gladia/solaria-36.86
25nvidia/parakeet-tdt-0.6b-v2 (fast-gpu-asr)6.9117254.19

Note: English tab, top 25 of 69. Mix of open-weight and proprietary models; RTFx blank where not reported. Multilingual and long-form tabs exist.

Source: Hugging Face Open ASR Leaderboard (hf-audio) · huggingface.co · results as of 2026-11-13 · retrieved 2026-11-17

Markdown · More Speech recognition (ASR)