Towards Automatic Methods to Detect Errors in Transcriptions of Speech Recordings

This work explores different methods to detect errors in transcriptionsof speech recordings. We artificially corrupt well transcribedspeech transcriptions with three types of errors: substitution, insertionand deletion on TIMIT phonemic transcriptions and WSJ wordtranscriptions. First, we use Bayesian model selection method bycomparing the log-likelihoods from alignment and phone recognizer,a final score is computed to make decision. In this method, weconsider two models, Bayesian Hidden Markov Model (HMM) anda Variational Auto-Encoder (VAE) combined with a HMM. Alternately,we build a biased ASR system with language models trainedon individual transcriptions, detection decision is based on Levenshteindistance (LD) between transcription and oracle path from decodedlattice. We evaluate the methods of detecting errors in corruptedTIMIT transcription, the best result (either using model selectionwith VAE model or biased ASR) achieves 7% equal errorrate on the Detection Error Tradeoff (DET) curve; we also evaluatethe methods of detecting errors in corrupted WSJ transcriptions, andthe best result (using biased ASR) achieves 3% equal error rate.