AI Briefing
KOSign in

AuroraASR-1M: A 949K-Parameter Speech Recognition Model Released Under Apache 2.0

·2026.10.18 21:48

Key point

The model achieves 11.43% WER on LibriSpeech test-clean with just 1.06 MB of INT8 weights.

Details

AuroraASR-1M is an extremely lightweight speech recognition model featuring 949K parameters and a weight-shared Conformer-CTC architecture. The model is released under the Apache 2.0 license.

Performance and Size

  • Model Size: Just 1.06 MB for INT8 weights.
  • Accuracy: Achieves 11.43% WER on LibriSpeech test-clean using greedy decoding without an external language model.
  • Speed: Runs at approximately 436× real-time on a single CPU thread.
open-sourceasrspeech-recognitionapache-2-0conformerlightweight-model
Read originalShare on Threads

This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.

Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.

Related

PyTorchKRVideoJev Releases System One Model for Video Segment Finding and QARedditTYGER Released: Open-Weights ASR with Full-Duplex Turn ControlDeepSeekDeepSeek Releases V4.1-Flash-1017 and Opens Search & Fetch APIsPyTorchKRStrata Engine Released to Run Qwen3.8-Flash-Next 125B on 12GB VRAMHuggingFace BlogIBM Releases Granite Switch 4.1 Models Combining Chat, Decisions, and Speech on Single EndpointGeekNewsSphinx: Open-Source Git Hook That Tests Developer Understanding of AI-Generated Code Before Deployment