Offline voice runs on ESP32

An ESP32 can now listen and speak without sending audio to a cloud service. Babytalk is a software library for ESP32-S3 and ESP32-P4 boards that turns speech into text and text into spoken audio. It supports C, MicroPython and AtomVM, and includes models, build details and performance measurements. The project says it’s designed for general-purpose English transcription, rather than recognizing only a preset list of commands. The project repository has the implementation and measurements.

For a business, the practical point is that voice input may work in a device that has no reliable internet connection, and audio can stay on the device. That could matter for a field tool, a radio with no keyboard, or voice control where Wi-Fi isn’t dependable. There are limits: the project calls for an ESP32-S3 or ESP32-P4 with 16 MB of flash and 8 MB of PSRAM, and says to skip it if you need very high accuracy or languages other than English. The speech-to-text model also takes 6 MB of flash in its 4-bit version, with text-to-speech requiring additional space.

This isn’t a ready-made business product. It’s a technical project with firmware and hardware requirements, so an operator would need to assess whether it fits the device and use case. Still, it’s the kind of development we build into an AI command centre: choosing where AI runs, and what data needs to leave the device.