Exploring recognition network representations for efficient speech inference on highly parallel platforms