Semi-Supervised Bootstrapping Approach For Neural Network Feature Extractor Training

This paper presents bootstrapping approach for training the Bottle-Neck neural networkfeature extractor which provides features  for subsequent GMM-HMM recognizer. One can use this recognizer to automatically transcribe the unsupervised data and assign the confidence of the transcription. Based on the confidence, segmentsare selected and mixed with supervised data and newNNs are trained. The automatic transcription can recover 40-55% in comparison to manually transcribed data. This is 3 to 5% absolute improvement over NN trainedon supervised data only. Using 70-85% of automaticallytranscribed segments with the highest confidence was foundoptimal to achieve this result. Dropping the rest of the data prevents training on low quality transcripts.