Benchmark Almanac

PazaBench v2

facebook/omniASR-LLM-7B-ZS 15.6% tops its leaderboard Whisper large-v3-turbo 73.3% leads the models we track
Not in model cards
Version 2

Models transcribe speech in 61 African languages. The leaderboard measures character error rate, word error rate and transcription speed.

Microsoft compares 53 models on 417,460 samples across 93 evaluation splits. The release adds four dataset groups.

Released
July 2026
Kind
Speech recognition and translation
Level
Everyday
Scored
Checked automatically
Answers
Open-ended
Tasks
All public
Measures
Capability
Language
Many languages
In and out
Audio → Text
Interaction
Single turn
Items from
Another benchmark
Run by
Anyone
Based on
PazaBench
Last model added
February 2025
Scores refreshed
November 20th, 2026
53
Entries on its own leaderboard, 4 models we track

facebook/omniASR-LLM-7B-ZS tops the leaderboard with 15.6% Whisper large-v3-turbo leads with 73.3% among models released by February 2025

Best score per entry from the benchmark's own leaderboard, read on November 20th, 2026. Where a model ran with several settings or harnesses, we keep its best. Names in gray aren't in our model list: fine-tunes, agent systems, products and older models. They don't count toward leaders. The best model we track is Whisper large-v3-turbo, at 73.3%. This view shows only models we track, each at its best.

The newest model with a score here came out in February 2025, and nobody has tested a newer one. Among the 20 highest-ranked models on the capability index, these have no score yet:

  1. 1 facebook/omniASR-LLM-7B-ZS 15.6%
  2. 2 facebook/omniASR-LLM-7B 16.1%
  3. 3 facebook/omniASR-LLM-3B 18.3%
  4. 4 facebook/omniASR-LLM-1B 19%
  5. 5 facebook/omniASR-CTC-7B 19.2%
  6. 6 facebook/omniASR-LLM-300M 19.3%
  7. 7 facebook/omniASR-CTC-3B 20.2%
  8. 8 facebook/omniASR-CTC-1B 20.6%
  9. 9 facebook/omniASR-CTC-300M 22.5%
  10. 10 facebook/mms-1b-all 30.1%
  11. 11 paza/microsoft-paza-mms-1b-all 33.9%
  12. 12 facebook/mms-1b-fl102 35.9%
  13. 13 facebook/wav2vec2-conformer-rel-pos-large-960h-ft 48.6%
  14. 14 facebook/hubert-large-ls960-ft 48.7%
  15. 15 facebook/hubert-xlarge-ls960-ft 49.3%
  16. 16 facebook/wav2vec2-large-960h-lv60-self 49.7%
  17. 17 paza/microsoft-paza-Phi-4-multimodal-instruct 51.6%
  18. 18 facebook/wav2vec2-large-robust-ft-libri-960h 52%
  19. 19 nvidia/parakeet-tdt-0.6b-v3 52.7%
  20. 20 facebook/data2vec-audio-large-960h 57%
  21. 21 facebook/wav2vec2-conformer-rope-large-960h-ft 57.9%
  22. 22 facebook/data2vec-audio-base-960h 60.6%
  23. 23 facebook/wav2vec2-large-960h 60.8%
  24. 24 paza/microsoft-paza-whisper-large-v3-turbo 63%
  25. 25 facebook/wav2vec2-base-960h 64.9%
  26. 26 nvidia/canary-qwen-2.5b 66.8%
  27. 27 openai/whisper-small.en 69.9%
  28. 28 openai/whisper-medium.en 70.4%
  29. 29 nvidia/canary-1b-v2 71%
  30. 30 Whisper large-v3-turbo 73.3%
  31. 31 openai/whisper-base.en 73.3%
  32. 32 Phi-4-multimodal 74.2%
  33. 33 efficient/speech-lite-whisper-large-v3-turbo-acc 74.3%
  34. 34 distil/whisper-distil-large-v3 74.3%
  35. 35 openai/whisper-large 75.9%
  36. 36 openai/whisper-tiny.en 76.2%
  37. 37 openai/whisper-large-v2 77.3%
  38. 38 distil/whisper-distil-medium.en 78.5%
  39. 39 kyutai/stt-2.6b-en 78.8%
  40. 40 distil/whisper-distil-large-v2 79.3%
  41. 41 Whisper large-v3 79.4%
  42. 42 efficient/speech-lite-whisper-large-v3-acc 80.3%
  43. 43 efficient/speech-lite-whisper-large-v3 82.4%
  44. 44 efficient/speech-lite-whisper-large-v3-turbo 82.8%
  45. 45 efficient/speech-lite-whisper-large-v3-fast 84.5%
  46. 46 efficient/speech-lite-whisper-large-v3-turbo-fast 89.9%
  47. 47 facebook/seamless-m4t-v2-large 92%
  48. 48 usefulsensors/moonshine-base 99.98%
  49. 49 Qwen2-Audio-7B-Instruct 135%
  50. 50 usefulsensors/moonshine-tiny 155.8%
  51. 51 ibm/granite-granite-speech-3.3-2b 292.8%
  52. 52 ibm/granite-granite-speech-3.3-8b 360.6%
  53. 53 Qwen/Qwen2-Audio-7B 501.3%

Source: Official PazaBench v2 leaderboard

Source: Official PazaBench v2 leaderboard

#EntryScoreSettingDateSource
1 facebook/omniASR-LLM-7B-ZS 15.6% Official leaderboard
2 facebook/omniASR-LLM-7B 16.1% Official leaderboard
3 facebook/omniASR-LLM-3B 18.3% Official leaderboard
4 facebook/omniASR-LLM-1B 19% Official leaderboard
5 facebook/omniASR-CTC-7B 19.2% Official leaderboard
6 facebook/omniASR-LLM-300M 19.3% Official leaderboard
7 facebook/omniASR-CTC-3B 20.2% Official leaderboard
8 facebook/omniASR-CTC-1B 20.6% Official leaderboard
9 facebook/omniASR-CTC-300M 22.5% Official leaderboard
10 facebook/mms-1b-all 30.1% Official leaderboard
11 paza/microsoft-paza-mms-1b-all 33.9% Official leaderboard
12 facebook/mms-1b-fl102 35.9% Official leaderboard
13 facebook/wav2vec2-conformer-rel-pos-large-960h-ft 48.6% Official leaderboard
14 facebook/hubert-large-ls960-ft 48.7% Official leaderboard
15 facebook/hubert-xlarge-ls960-ft 49.3% Official leaderboard
16 facebook/wav2vec2-large-960h-lv60-self 49.7% Official leaderboard
17 paza/microsoft-paza-Phi-4-multimodal-instruct 51.6% Official leaderboard
18 facebook/wav2vec2-large-robust-ft-libri-960h 52% Official leaderboard
19 nvidia/parakeet-tdt-0.6b-v3 52.7% Official leaderboard
20 facebook/data2vec-audio-large-960h 57% Official leaderboard
21 facebook/wav2vec2-conformer-rope-large-960h-ft 57.9% Official leaderboard
22 facebook/data2vec-audio-base-960h 60.6% Official leaderboard
23 facebook/wav2vec2-large-960h 60.8% Official leaderboard
24 paza/microsoft-paza-whisper-large-v3-turbo 63% Official leaderboard
25 facebook/wav2vec2-base-960h 64.9% Official leaderboard
26 nvidia/canary-qwen-2.5b 66.8% Official leaderboard
27 openai/whisper-small.en 69.9% Official leaderboard
28 openai/whisper-medium.en 70.4% Official leaderboard
29 nvidia/canary-1b-v2 71% Official leaderboard
30 Whisper large-v3-turbo 73.3% Oct 2024 Official leaderboard
31 openai/whisper-base.en 73.3% Official leaderboard
32 Phi-4-multimodal 74.2% Feb 2025 Official leaderboard
33 efficient/speech-lite-whisper-large-v3-turbo-acc 74.3% Official leaderboard
34 distil/whisper-distil-large-v3 74.3% Official leaderboard
35 openai/whisper-large 75.9% Official leaderboard
36 openai/whisper-tiny.en 76.2% Official leaderboard
37 openai/whisper-large-v2 77.3% Official leaderboard
38 distil/whisper-distil-medium.en 78.5% Official leaderboard
39 kyutai/stt-2.6b-en 78.8% Official leaderboard
40 distil/whisper-distil-large-v2 79.3% Official leaderboard
41 Whisper large-v3 79.4% Nov 2023 Official leaderboard
42 efficient/speech-lite-whisper-large-v3-acc 80.3% Official leaderboard
43 efficient/speech-lite-whisper-large-v3 82.4% Official leaderboard
44 efficient/speech-lite-whisper-large-v3-turbo 82.8% Official leaderboard
45 efficient/speech-lite-whisper-large-v3-fast 84.5% Official leaderboard
46 efficient/speech-lite-whisper-large-v3-turbo-fast 89.9% Official leaderboard
47 facebook/seamless-m4t-v2-large 92% Official leaderboard
48 usefulsensors/moonshine-base 99.98% Official leaderboard
49 Qwen2-Audio-7B-Instruct 135% Aug 2024 Official leaderboard
50 usefulsensors/moonshine-tiny 155.8% Official leaderboard
51 ibm/granite-granite-speech-3.3-2b 292.8% Official leaderboard
52 ibm/granite-granite-speech-3.3-8b 360.6% Official leaderboard
53 Qwen/Qwen2-Audio-7B 501.3% Official leaderboard
#ModelScoreSettingDateSource
1 Whisper large-v3-turbo 73.3% Oct 2024 Official leaderboard
2 Phi-4-multimodal 74.2% Feb 2025 Official leaderboard
3 Whisper large-v3 79.4% Nov 2023 Official leaderboard
4 Qwen2-Audio-7B-Instruct 135% Aug 2024 Official leaderboard

Rankings by part

We read 2 parts from the benchmark’s own leaderboard: Word error rate and Inverse real-time factor. Each ranks models on one part of the test. Whisper large-v3-turbo leads both. Arrows show how far a model moved from its overall rank among the models we track.

Word error rate

4 models we track have a score here.

  1. 1 Whisper large-v3-turbo 96.7%
  2. 2 Whisper large-v3 ↑1 98.1%
  3. 3 Phi-4-multimodal ↓1 112.8%
  4. 4 Qwen2-Audio-7B-Instruct 160%

Number one over time

Lab claims count from the day the card came out, independent runs from the model’s release.

OpenAI
Own leaderboard Not tested since Jul 2026

Source: Official PazaBench v2 leaderboard

Scores over time

All the scores we have, one dot each, so a model tested at several settings appears more than once.

Plot with other benchmarks
On the benchmark’s own leaderboard Best so far
0% 20% 40% 60% 80% 100% 2024 2025 Whisper large-v3-turbo Whisper large-v3 0% 20% 40% 60% 80% 100% 2024 2025 Whisper large-v3-turbo

Source: Official PazaBench v2 leaderboard

No lab model card reports it yet

We read the headline results of 540 launch posts from 57 labs, and none of them list this benchmark.

Versions

More in Audio and video

All 695 in Audio and video
  • Video-MMMU In use 300 expert-level lecture videos with questions testing perception, comprehension and adaptation of knowledge. Accuracy. Qwen 3.8 Max 88.7%
  • LVBench In use Hours-long video understanding with multiple-choice questions on reasoning, grounding and summarization. Gemini 4 Argon 91.7%
  • MMVU In use 3,000 expert-annotated questions on videos across 27 disciplines. Multiple-choice and open-ended accuracy. Qwen 3.8 Max 82.4%
  • MLVU In use Multi-task long video understanding (minutes to hours) with multiple-choice and generation tasks. Accuracy. Qwen 3.8 Max 90.8%
  • FLEURS In use Models transcribe read speech in 102 languages built from FLoRes-101 sentences. Model cards usually report word error rate, so lower scores are better. Phi-4-multimodal 4.00%
  • VoiceBench In use Spoken instructions with speaker, environment and content variation. Voice assistant responses scored by GPT judge or rules. Qwen3.5-Omni-Plus 93.1%

Data audit

8 of 9 checks pass, 1 warning. Look first at the Newest score check. The newest came 21 months ago, in February 2025, for Phi-4-multimodal. Every benchmark

Catalog record 2 pass
  • Pass Task access All tasks are public.
  • Pass Confidence Checked against a primary source.
Scores 1 warning 2 pass 2 notes
  • Pass Independent runs Scores from the official leaderboard (53 entries).
  • Pass Models scored 4 models we track, plus 49 other entries on its leaderboard.
  • Warning Newest score The newest came 21 months ago, in February 2025, for Phi-4-multimodal.
  • Note Frontier models None of its models are on the capabilities index.
  • Note Headroom Lower is better, so there’s no ceiling to reach.
Leaderboard import 3 pass
  • Pass Importer We read the board on huggingface.co: 4 rows for models we track and 49 other entries at the last read.
  • Pass Last read November 20th, 2026.
  • Pass Read worked The last read worked.
Consistency 1 pass 1 note
  • Note Lab claims No model has both a lab number and an independent run.
  • Pass Known issues No published audit found problems.

How we check