Research · Oct 31, 2026

Seven speech models on real news broadcasts: the leaderboard #1 ties a 0.6B model

Result

No significant difference: 7 models within 0.94 WER points (smallest Holm-corrected p = 0.65). Speed differs 23x.

Hardware: NVIDIA H100 80GB (cloud, spot)
Data: 31 broadcasts · 2.6 h
Run: 72 min · ~$3–5
Metric: WER, 95% CI

Parakeet Unified EN 0.6B7.34

Qwen3-ASR 1.7B7.58

Whisper large-v3-turbo7.64

Parakeet TDT 0.6B v37.97

Nemotron-3.5-Transcribe 30B-A4B8.00

Granite Speech 5.0 2B8.02

freetranscribe parakeet-ft-v18.28

Word error rate (%): 5–11 · dot = measured · bar = 95% CI · overlapping, so not significant

The Open ASR Leaderboard has a clear #1 this month: a 30B mixture-of-experts model. Below it sit models a tenth its size. So, a simple question: point all of them at real news broadcasts (studio anchors, field reporters, press conferences, crowd noise). Does the order hold? It doesn't. On one rented H100 the seven land within one WER point of each other, and a paired test can't tell any two apart. What does differ is cost.

Read the full study on Every Local AI