Multi-Frame Cross-Entropy Training for Convolutional Neural Networks in Speech Recognition
This work addresses speech recognition accuracy for users of CNN acoustic models, offering a novel training approach with competitive gains.
The paper tackles the problem of improving speech recognition accuracy by proposing Multi-Frame Cross-Entropy training for CNNs, which increases label information from multiple adjacent frames with minimal computational cost, resulting in large WER improvements on hub5 and rt02 datasets after training on the 2000-hour Switchboard benchmark.
We introduce Multi-Frame Cross-Entropy training (MFCE) for convolutional neural network acoustic models. Recognizing that similar to RNNs, CNNs are in nature sequence models that take variable length inputs, we propose to take as input to the CNN a part of an utterance long enough that multiple labels are predicted at once, therefore getting cross-entropy loss signal from multiple adjacent frames. This increases the amount of label information drastically for small marginal computational cost. We show large WER improvements on hub5 and rt02 after training on the 2000-hour Switchboard benchmark.