I assume it would be possible to make yourself with the OpenAI api together with a locally run voice model to only detect the activation word. There might be of the shelf solutions for this, but I am not aware of any.