This code is the Python implementation for the method blank collapse proposed in the paper Blank Collapse: Compressing CTC emission for the faster decoding accepted to Interspeech 2023.
This repository includes the function for blank collapse and its jupyter test code using torchaudio.
- The following figure shows the resulting image from blank collapse method whose total length is shrunk by collapsing unnecessary blank frames.
- This table shows how much this method can save the time in beam search decoding.
