Blog / Engineering
Engineering posts from the OnType blog
From Spoken Words to Clean Text: How Inverse Text Normalization Works
When you say "three thousand dollars", you expect to see "$3,000" — not the literal words. This is inverse text normalization, and we built a custom engine from scratch to make it work across languages.
Engineering·
Building On-Device Speech Recognition with MLX on Apple Silicon
How we built mlx-swift-asr — an open-source, native Swift library that runs Qwen3-ASR on Apple Silicon at 4–6x real-time speed with zero cloud dependency.
Engineering·