Unsupervised acoustic model training using multiple seed ASR systems

Unsupervised acoustic modeling can offer a cost and time effective way of creating a solid acoustic model for any under-resourced language. This paper explores the novel idea of using two independent ASR systems to transcribe new speech data, align and filter the ASR hypotheses and use the presumably correct transcriptions to iteratively improve the two seed ASR systems. In parallel, the newly transcribed speech is used to retrain the mainstream ASR system. The methodology leads to WER relative improvements of 5.5% after the first iteration. The experiments are made with data in the Romanian language. Index Terms unsupervised acoustic modeling, speech recognition, unsupervised training, under-resourced languages