Give Home Assistant a wake phrase of your own

You can give a home voice assistant an activation phrase without first recording yourself saying it hundreds of times. openWakeWord trains a small detector using generated speech and examples without the target phrase.[3]

This is a documentation-based procedure, checked on 20 October 2026, with no training or hardware trial performed. Start with current Home Assistant OS, a working Assist pipeline and an ATOM Echo already configured through Home Assistant’s device tutorial.[4]

Train the phrase

Choose an uncommon English phrase of three or four syllables. Open the training environment linked from Home Assistant’s Wake words for Assist guide. Enter it in target_word and run that section. Listen to the generated pronunciation; adjust the spelling and rerun if necessary.[1]

Choose Runtime → Run all and leave the tab open. Colab resource availability can delay training. Save the downloaded .tflite; this procedure does not use the accompanying .onnx file.[1]

Install it on the server

In Home Assistant, install and start openWakeWord under Settings → Apps. Configure its discovered Wyoming integration under Settings → Devices & services. Using the Samba app to reach the server, copy the model into /share/openwakeword, creating that folder if needed. Reload the openWakeWord Wyoming integration after copying it.[2]

Edit the intended assistant under Settings → Voice assistants. Open its three-dot menu and choose Add streaming wake word; select openwakeword and your model, then save. This file is for the streaming server engine. It will not appear in the ESPHome selector for on-device microWakeWord models.[1]

Under Settings → Devices & services → ESPHome, open the ATOM Echo device. Enable Use wake word and select the assistant you just edited.[1]

Check the phrase where you will use it

Say the phrase, wait for the Echo’s blue light, then give a command that already works in your setup. The light shows the activation stage; completing the command also exercises the rest of the Assist pipeline.[4]

Try several normal speaking positions, and listen for activations during ordinary conversation. Keep separate notes for missed invocations and unwanted triggers. A model that responds when spoken to has passed only half of that practical check.

The app’s threshold controls how readily audio activates the detector; a higher value means fewer activations. Raising trigger_level requires more activations before registering a detection. Change one setting at a time and repeat the same phrases, so an apparent improvement is easier to interpret.[2]

Keep the model that works best in your room, along with the phrase and settings used to train it. A different name is easy to choose; your comparison tells you whether the change helped.

Report an error or suggest a correction