Created by: xinghai-sun

• Create a simplified model configuration for DeepSpeech2 on V2 APIs.
  • With fixed-length (by padding) batches.
  • With only bidirectional-GRU.
  • With only greedy decoder (max + post-processing).
• Here "simplified" refers to the above limitations, so that we can move on without being blocked by other un-resolved tasks (e.g. 2228, 2230 etc.).
• About network architecture please refer to DS2 paper and the design doc.
• Please test it on GPUs, and make sure to achieve at least a normal convergence.
• Please pull your codes and docs into PaddlePaddle/models.