MultiReader Technique which allows domain adaptation training a more accurate model that supports multiple keywords (i.e., “OK Google” and “Hey Google”) as well as multiple languages/dialects