The A.I. Daily
AuroraASR-1M released: sub-1M-parameter lightweight speech recognition model
AuroraASR-1M uses a weight-shared Conformer-CTC architecture with 949K parameters and 1.06MB INT8 weights, running about 436x real-time on one CPU thread. It reaches 11.43% WER on LibriSpeech test-clean with greedy decoding and no external language model, under Apache 2.0.