DLNLP学习笔记06(Speech Recognition: Neural Transducer模型&MoChA模型&总结)

1 Neural Transducer:根据一个window size将多个输入进行attention之后,再输入到decoder。实际实验结果,加了attention之后window size大小对模型效果影响不大。

DLNLP学习笔记06(Speech Recognition: Neural Transducer模型&MoChA模型&总结)


2 MoChA (Monotonic Chunkwise Attention):动态地移动window。

DLNLP学习笔记06(Speech Recognition: Neural Transducer模型&MoChA模型&总结)

DLNLP学习笔记06(Speech Recognition: Neural Transducer模型&MoChA模型&总结)


3 总结:

DLNLP学习笔记06(Speech Recognition: Neural Transducer模型&MoChA模型&总结)