Speech Recognition

• Recurrent Neural Network Based Language Model
  • 50% reduction of Perplexity
  • mixture of several Basic RNN Architectures
  • Wall Street Journal task
  • connectionist language models are superior to standard n gram techniques, except their high computational (training) complexity
  • break the myth that language modeling is just about counting n-grams, and that the only reasonable way how to improve results is by acquiring new training dat
• Towards End-To-End Speech Recognition with Recurrent Neural Networks
  • character-level speech recognition system that directly transcribes audio data with text using a recurrent neural network
  • combination of the deep bidirectional LSTM recurrent neural network architecture and a modified Connectionist Temporal Classification (CTC) objective function
  • word error rate
  • Wall Street Journal task