Speech-Recognition

based on NeMo

• 

based on espnet

• no batch inference yet?
• 

based on deepspeech.pytorch

PyTorch implementation of DeepSpeech2 trained with the CTC objective.

differences to deepspeech.pytorch

• no use of warp-ctc, instead torch.nn.CTCLoss
• powered by pytorch-lightning

results

• after 8 epochs and 24hours with Adam

python evaluation.py --model epoch=8.ckpt --datasets test-clean
2528 of 2620 samples are suitable for training
100%|█████████████████████████████████████| 127/127 [02:12<00:00,  1.04s/it]
Test Summary    Average WER 9.925       Average CER 3.239

python evaluation.py --model epoch=8.ckpt --datasets test-other
2893 of 2939 samples are suitable for training
100%|███████████████████████████████████████| 145/145 [01:19<00:00,  1.83it/s]
Test Summary    Average WER 27.879      Average CER 11.739

Datasets

Librispeech

1. to download data see: https://github.com/dertilo/speech-to-text/corpora/download_corpora.py

• splits

  datasets = [
      ("train", ["train-clean-100", "train-clean-360", "train-other-500"]),
      ("eval", ["dev-clean", "dev-other"]),
      ("test", ["test-clean", "test-other"]),
  ]
  

• number of samples

  train got 281241 samples
  eval got 5567 samples
  test got 5559 samples