The recognition accuracy of on-device models is typically lower than that of large cloud models, especially in noisy environments or scenarios with many technical terms.