Hayamimi runs real-time speech recognition on CPU only, hits 3.8% error

The release follows a pattern where CPU-based speech recognition has often depended on a single model. Hayamimi takes a different approach: it detects the language of each utterance and routes it to a dedicated model, which allows it to handle multiple languages without a GPU. The system uses INT8-quantized ONNX models running on sherpa-onnx, so PyTorch and CUDA are not required. On the accuracy side, a two-pass correction re-decodes the previous utterance after a 2-second silence, improving the Japanese Character Error Rate from 15.5% to 12.0%. The project also supports speaker labeling, hotwords, translation, and a browser dashboard through a local HTTP server. These features suggest the developer is targeting live production use — streaming overlays, transcription pages, and network audio input — rather than just lab experiments. The documented limitations, such as not handling mixed-language utterances or separating simultaneous speakers, indicate the system is designed for clear, single-speaker scenarios rather than complex meeting environments.