Radar ·
AuroraASR-1M: 1MB English speech recognition model for CPU
An English speech model with 949 thousand parameters for CPU, reported by the author at about 436 times real time and 11.43% WER on test-clean.
AuroraASR-1M is described as a Conformer-CTC model with 949 thousand parameters (1.06 MB in INT8, Apache 2.0 license), trained on the LibriSpeech dataset for English speech recognition. The author reports about 436 times real time on a single CPU thread, but with a word error rate (WER) of 11.43% on test-clean, against 5.6% for Whisper Tiny. The author's stated use is offline, CPU-only work where model size and latency matter more than accuracy.
The case illustrates the trade-off between speed, size and quality in small speech models. To verify it, consult the original page indicated by the author and check the test set, the WER metric and the conditions of the speed measurement before drawing any conclusion. Rota Nacional does not offer this model nor guarantee similar performance; what the platform offers for audio is transcription, which already passes through the same personal data detection barrier.
Original source: AuroraASR-1M: modelo de reconhecimento de fala em inglês de 1MB para CPU