Deep learning for speech algorithms (recognition, enhancement, echo cancellation, synthesis, etc) on low power devices.