45% fewer errors than the nextbest published system.
Composite WER of 3.10% across 8 standard OpenASR benchmarks for theUniversal tier, and a macro WER of 24.6% across 16 Indian languages and7 datasets - four of which had no prior ASR coverage at all.
| Capability | Universal tier | Indian Languages tier |
|---|---|---|
| Supported languages | 204 (~96.8% global population) | 55 (16 on public benchmarks, 4 first-ever) |
| Composite / Macro WER | 3.10% | 24.6% macro / 11.9% Hindi |
| Relative gain vs best baseline | 45% (OpenASR) | 14% vs IndicWhisper, 40% vs NVIDIA Conformer-CTC Large |
| Latency (xRT) | 0.0068 (~146× real time, A100) | Comparable, confirming on target hardware |
| Streaming | Yes, ~200ms partials | Yes, same engine wrappers |
| Speaker diarization | Yes - in production | Yes - same diarization layer |
WER by benchmark - Universal ASR vs. NVIDIA Canary-Qwen 2.5B
Lower is better. Source: OpenASR leaderboard, October 2025.
LibriSpeech Clean
audiobooks, clear speech
0.71%
1.42%
SPGISpeech
financial earnings calls
1.1%
1.9%
TedLium
ted talks
1.43%
2.71%
LibriSpeech Other
audiobooks, noisy audio
2.17%
2.87%
AMI
meetings
4.19%
9.12%
VoxPopuli
parliament speech
4.34%
5.63%
GigaSpeech
mixed web audio
4.99%
9.43%
Earnings22
earnings calls
5.83%
9.53%
Composite across all 8 benchmarks: 3.10% for Shunya vs. 5.63% for Canary-Qwen 2.5B - a 45% relative reduction.
| System | Languages | Composite WER |
|---|---|---|
| Shunya Universal ASR | 204 | 3.10% |
| OpenAI Whisper Large v3 | 99 | 5.40% |
| NVIDIA Canary-Qwen 2.5B | 3 | 5.63% |