Semi-supervised training on 17.55 hours of untranscribed Somali speech yields 7.74% relative WER reduction and 6.55% LM perplexity improvement over a 1.57-hour baseline using TDNN-F models and confidence filtering.
As we only have less than two hours of transcribed Somali acoustic data, increasing the pool of in-domain data by semi-supervised training was an attractive option
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CL 1years
2019 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
Improved low-resource Somali speech recognition by semi-supervised acoustic and language model training
Semi-supervised training on 17.55 hours of untranscribed Somali speech yields 7.74% relative WER reduction and 6.55% LM perplexity improvement over a 1.57-hour baseline using TDNN-F models and confidence filtering.