3/ It also processes audio in slices as small as 160ms. Before anyone quotes that wrong: that's the streaming step, not end-to-end latency. Different numbers, different meaning. What it buys you in practice is that understanding starts while the person is still mid-sentence,