Who am I to argue with your experience? I can only offer the current literature on the topic [0]. (They report a 5.1% error) > These systems typically use deep

Who am I to argue with your experience? I can only offer the current literature on the topic [0]. (They report a 5.1% error) > These systems typically use deep convolutional neural network (CNN) architectures ... driving the word error rate on the benchmark Switchboard corpus down from its mid-2000s plateau of around 15% to well below 10%. [0]: https://arxiv.org/pdf/1708.06073.pdf https://arxiv.org/pdf/1708.06073.pdf