Transcripts in Automatic Speech Recognition systems are commonly generated using only an acoustic model, typically an “end-to-end” CTC-based network which matches audio and text without additional alignment information. However, ambiguities in the transcription, for example when collapsing repeated characters and removing blanks, can exist as the CTC-based network has little prior linguistic knowledge. This is where a language model comes in, as it can improve performance by helping solve those decoding ambiguities.
We provide a trained language model on Wall Street Journal data. The language model we use is based on prefix beam search KenLM which imposes a language model constraint on the new predicted character based on previous (most probable) prefixes. Specifically, the model we trained is the Baidu’s CTC decoder with N-Gram LM implementation using 6 as N and train on WSJ data.