AuroraASR-1M: A 1MB Speech Recognition Model Running 436x Real-Time on One CPU Thread
Tall_Abrocoma_3533 · reddit · 2026-10-18
AuroraAI-Research released AuroraASR-1M, an extremely lightweight ASR model: 949K parameters with a weight-shared Conformer-CTC architecture, just 1.06MB of INT8 weights, 436× real-time on a single CPU thread, and 11.43% WER on LibriSpeech test-clean with greedy decoding and no external LM. Apache 2.0 licensed.
More from Research
- AI predicted 200 million protein structures and gave them away, skipping drug discovery's slowest step — rand_longevity · 2026-10-18
- KAIST and DeepMind propose Declarative Attention: models declare which context they need, cutting KV reads by 46% — coallaoh · 2026-10-18
- Fields Medalist Gowers warns AI could erode future mathematicians' motivation — IgorCarron · 2026-10-18
- SearchJev: a fast System-1 model for search agents with 5.3x faster decisions and 45%→54% accuracy — CShorten30 · 2026-10-18
- Astraea cognitive architecture: 17 days of fixes in timing, concurrency and memory — Financial_Tadpole121 · 2026-10-18
- 100,000 submissions in a month: comment urges putting Lean in the loop earlier — IgorCarron · 2026-10-18