Yehor/kenlm-uk

Community

• Discord: https://discord.gg/yVAjkBgmt4
• Speech Recognition: https://t.me/speech_recognition_uk
• Speech Synthesis: https://t.me/speech_synthesis_uk
• Natural Language Processing: https://t.me/nlp_uk

Overview

Different KenLM models for Ukrainian.

Metrics

Tested with an acoustic model from w2v-xls-r-uk model:

Model	CER	WER
no LM	0.0412	0.2206
lm-3gram-10k (alpha=0.1)	0.0398	0.2191
lm-4gram-10k (alpha=0.1)	0.0398	0.219
lm-5gram-10k (alpha=0.1)	0.0398	0.219
lm-3gram-30k	0.038	0.2023
lm-4gram-30k	0.0379	0.2018
lm-5gram-30k	0.0379	0.202
lm-3gram-50k	0.0348	0.1826
lm-4gram-50k	0.0347	0.1818
lm-5gram-50k	0.0347	0.1821
lm-3gram-100k	0.031	0.1588
lm-4gram-100k	0.0308	0.1579
lm-5gram-100k	0.0308	0.1579
lm-3gram-300k	0.0261	0.1294
lm-4gram-300k	0.0261	0.1293
lm-5gram-300k	0.0261	0.1293
lm-3gram-500k	0.0248	0.1209
lm-4gram-500k	0.0247	0.1207
lm-5gram-500k	0.0247	0.1209

Files of the KenLM models are under the Files and versions section.

Attribution

• Chaplynskyi, D. et al. (2021) lang-uk Ukrainian Ubercorpus [Data set]. https://lang.org.ua/uk/corpora/#anchor4