Experiment Code for a Theory of Unsupervised Speech Recognition

This repository contains the source code for the paper "A Theory of Unsupervised Speech Recognition": https://www.researchgate.net/publication/370605684_A_Theory_of_Unsupervised_Speech_Recognition.

Dependencies

• fairseq >= 1.0.0 with dependencies for wav2vec-u

How to run it?

0. Change $CONDA_ROOT, $FAIRSEQ_ROOT, $KALDI_ROOT and $KENLM_ROOT in run_{synthetic, analysis}.sh to those of your own.

1. For phase transition experiments: bash run_synthetic.sh ${l1,mmd,jsd,wasserstein} ${circulant,debruijn,hypercube}

2. For further analysis such as the effect of training with discriminator reset, discriminator type and generator type:

bash run_analysis.sh ${gan_type} ${graph_name} ${gen_type} ${discrim_type} 2 2  # The effect of discriminator reset
bash run_analysis.sh ${gan_type} ${graph_name} ${gen_type} ${discrim_type} 4 4  # The effect of discriminator type
bash run_analysis.sh ${gan_type} ${graph_name} ${gen_type} ${discrim_type} 6 6  # The effect of generator type

3. To generate figures from the paper, please check out synthetic_asr_u.ipynb for more details. Currently, figures for the GAN-based experiments are generated based on manually-created .csv files.