hayamimi (早耳) is a real-time, multilingual speech-to-text engine that runs entirely on CPU, routing each spoken utterance to whichever specialist ONNX model best fits its language via sherpa-onnx instead of relying on one general-purpose model. It natively covers Japanese, Mandarin, Korean, Cantonese, English, and 24 European languages, and falls back to Meta's Omnilingual ASR for roughly 1,600 additional languages. Partial subtitles stream in every ~0.5s while you're still talking, a finalized line lands about 100ms after you stop, and a two-pass batch re-decode after a silence gap further sharpens the transcript. Built-in speaker turn-taking labels, live translation, and a browser dashboard with an OBS browser-source overlay make it usable for streaming captions and meeting transcripts without a GPU or cloud API. On measured real broadcast Japanese audio it reaches 5.8% CER versus 13.8% for whisper-large-v3-turbo on the same clips, while running 9-16x faster than real time on CPU alone.