CTCDecoder

Why we included this project

The decoder is the part of a CTC pipeline that often gets glossed over until the output comes out wrong. This small Python package implements the standard options, best path and beam search for the common cases, plus lexicon search, prefix search, and token passing for when you need more than a greedy guess. The beam search decoder can optionally use a bigram language model built from your own text, and the lexicon search relies on a BK-tree to return dictionary words within a chosen edit distance. It is a focused, dependency-light library rather than a full training framework, so it is easy to drop into an existing pipeline or to read through if you want to understand how each algorithm works. If you build speech-to-text or handwriting recognition on a CTC-trained network, this is a handy reference and a solid baseline to measure your own decoder against.