A low-cost ESP32 speaker with wake word on device and conversation in the cloud.
A voice product built on a deliberately inexpensive MCU. Wake-word detection and audio front-end processing run locally; streaming speech recognition, the language model and speech synthesis run behind an AI gateway.
Conversational latency had to feel immediate on a part with a few hundred kilobytes of usable RAM, over consumer Wi-Fi, without a per-unit cost that broke the product.
We streamed audio in both directions and started playback on the first synthesized chunk rather than waiting for a complete response. Local wake-word and echo cancellation kept the radio quiet until it mattered.
An ESP32-S3 with a dual-microphone array, I2S codec and amplifier on a compact 4-layer board.
On-device wake word and voice activity detection, with streaming ASR, an LLM and TTS reached through a gateway.
ESP-IDF firmware with a WebSocket session protocol, provisioning flow and OTA, behind a stateless session service.
Hardware
AI
Embedded
Cloud