Show HN: Live speech-to-text on an ESP32-P4 microcontroller, no cloud (github.com)

A new project demonstrates live English speech recognition on the ESP32-P4 microcontroller without relying on cloud services. This achievement is significant for the AI/ML community as it showcases the capabilities of AI running directly on low-power hardware, utilizing two RISC-V cores and innovative PIE SIMD instructions. The system uses the Kroko model, a 65.7 million-parameter streaming Zipformer2 transducer, which has been optimized to fit into 26.65 MB of weight files, enabling real-time capabilities with a word error rate (WER) of 4.34% on clean test data. Technical details reveal that the speech recognition runs at 0.70× real-time, processes audio in chunks, and maintains a streamlined interaction where the final word appears on screen approximately two seconds after being spoken. The setup utilizes a microSD card for weight storage, with the firmware demonstrating advanced voice control capabilities across six interactive demos. This innovation not only deepens the potential for embedded AI applications but also pushes the boundaries of what's achievable in edge computing, marking a promising step toward more efficient, standalone AI systems.