GitHub radar
Hayamimi: Real-Time Multilingual Subtitles on CPU, No Cloud
Realtime speech to text using Python on CPU without any cloud API or GPU.
Hayamimi (which means ‘quick ear’ in Japanese) is a python based tool that provides realtime multilingual speech to text transcription using CPU only, without using any cloud API or GPU. It uses quantized (INT8) ONNX models to redirect utterances to appropriate models based on their language. Each of Japanese, Chinese, Korean, Cantonese, English, 24+ EU languages has their own specialized model, while others are redirected to Meta’s Omnilingual ASR. Achieved 5.8% character error rate on real broadcast Japanese audio, which is less than half compared to Whisper large-v3-turbo (13.8%), at realtime 10–50x. Subtitles are partially displayed during your speech (~0.5 sec interval), finalized subtitles (~100ms delay) are displayed when you finish. Live subtitles are displayed on web dashboard. Speaker labels (e.g., S1, S2) and translations are available too. OBS Overlay for streaming caption is also supported. Low memory usage (< 2GB). Python 3.10 or higher, ffmpeg, ~3.1GB of downloadable models (1.1GB for minimal japanese+english installation)
Why a vibe-coder should care
Do you want to have realtime speech to text transcription in meetings, calls, and streaming, without using any cloud API, and without using GPU? Try Hayamimi, and use what you have (i.e., laptop). Higher accuracy than Whisper (popular alternative) and multilingual. No subscription necessary.
How to install
Copy this and send it to your agent — Claude Code, Codex, any of them:
Install hayamimi from https://github.com/oboroge0/hayamimi following the README — create a virtual environment, install dependencies, download models — and start real-time transcription from the microphone with the dashboard at http://localhost:8833/dashboard. Ask me if you need system microphone permissions.
Runs on a regular laptop — Python and ffmpeg required; downloads ~3 GB of models on first run (or ~1 GB in minimal mode for Japanese and English only).
Open on GitHub▌ More finds