boknilev / asr-repr-analysis

Analyzing ASR Representations

This repository contains code for our paper on analyzing speech representations in end-to-end automatic speech recognition models:

"Analyzing Hidden Representations in End-to-End Automatic Speech Recognition Systems", Yonatan Belinkov and James Glass, NIPS 2017.

Requirements

• lmdb
• tds
• nngraph

Instructions

1. First prepare a dataset in LMDB format according to the instructions in deepspeech.torch. We provide a custom MakeLMDBTimes.lua file to process a dataset with time segmentation such as TIMIT.
2. Run train.lua with the following arguments:

• loadPath: DeepSpeech-2 model trained with deepspeech.torch
• trainingSetLMDBPath, validationSetLMDBPath, testSetLMDBPath: top folders for the LMDB training/validation/test sets
• reprLayer: representation layer name (input, cnn1, cnn2, rnn1, rnn2, etc.)
• predFile: file to save predictions

See train.lua for more options, such as controlling convolution strides, using a window of features around the frame or predicting phone classes.

Citing

If you use this code, please consider citing our paper:

@InProceedings{belinkov:2017:nips,
  author     = {Belinkov, Yonatan and Glass, James},
  title      = {Analyzing Hidden Representations in End-to-End Automatic Speech Recognition Systems},
  booktitle  = {Advances in Neural Information Processing Systems (NIPS)},
  month      = {December},
  year       = {2017}
}

Acknowledgements

This project uses code from deepspeech.torch.