Generate KenLM trie based language model

[kdavis-mozilla]

No description provided.

[kdavis-mozilla]

Generated trie based language models using separately the data sets LibriSpeech, VoxForge, and English Wikipedia.

Instead of creating a single trie model the parameters defining the model were varied. For example the n-gram depth, pruning, vocab estimate, array, quantization, bit depth...

Through all of these variations the tries produced by the English Wikipedia text, an 11GB text, were always greater than the current 1.48 GB language model. So they would not result in a size win. Thus were not further pursued.

For the LibriSpeech and VoxForge tests were made against the librivox test clean all with the same acoustic model to determine which language model performed best. Results are here[1].

[lock]

This thread has been automatically locked since there has not been any recent activity after it was closed. Please open a new issue for related bugs.