Repository files navigation

Useful functions for E2E Speech Recognition training with PyTorch and CUDA.

Here is a simple use case with Reinforcement Learning and RNN-T loss:

blank=torch.tensor([0],dtype=torch.int).cuda()space=torch.tensor([1],dtype=torch.int).cuda()xs=model.greedy_decode(xs,sampled=True)torch_edit_distance.remove_blank(xs,xn,blank)rewards=1-torch_edit_distance.compute_wer(xs,ys,xn,yn,blank,space)nll=rnnt_loss(zs,ys,xn,yn)loss=nll*rewards

Levenshtein edit-distance with detailed statistics for ins/del/sub operations.

Merge repeated tokens, useful for CTC-based model.

Remove unnecessary blank tokens, useful for CTC, RNN-T, RNA models.

Remove leading, trailing and repeated middle separators.

• C++11 compiler (tested with GCC 9.4.0).
• Python: 3.5, 3.6, 3.7, 3.8, 3.9 (tested with version 3.8).
• PyTorch >= 1.5.0 (tested with version 1.13.1+cu116).
• CUDA Toolkit (tested with version 11.2).

There is no compiled version of the package. The following setup instructions compile the package from the source code locally.

pip install torch_edit_distance

git clone https://github.com/1ytic/pytorch-edit-distancecd pytorch-edit-distancepython setup.py install

python -m torch_edit_distance.test