CTCWordBeamSearch

Why we included this project

Many CTC-based recognizers default to greedy decoding or a plain beam search, which can produce misspelled or out-of-vocabulary text. This decoder instead restricts output to words from a dictionary you supply and ranks candidate sequences with a language model trained on your own corpus, so the final text stays within a known vocabulary. That fits speech transcription and handwritten form reading where the valid word set is already defined. The Python package installs and passes its tests without fuss, and the README walks through a minimal example that shows how to feed a model's probability matrix and get decoded strings back. If you are building a recognizer and want to avoid post-processing guesswork, this is a straightforward component to add.