LibriSpeech test-clean, 500 utterances. 1.7B is borderline real-time on M5 (RTF 0.944). 0.6B (3.30% WER, 0.263 RTF) is the practical choice for MacBook.
LibriSpeech test-clean, 500 utterances, per-utterance simul-streaming. AlignAtt border detection with 20 alignment heads. Platform: Apple M5 32GB (MLX fp16). benchmark_mlx_simul.py: reusable benchmark script for MLX backends.