Are Some Speech Recognition Errors Easier to Detect than Others?

This study investigates whether some speech recognition (SR) errors are easier to detect and what patterns can be identified from those errors. Specifically, SR errors were examined from both nonlinguistic and linguistic perspectives. The analyses of non-linguistic properties revealed that high error ratios and consecutive errors lowered the ease of error detection. The analyses of linguistic properties showed that ease of error detection was associated with changing parts-of-speech of reference words in SR errors. Additionally, syntactic relations themselves and the change of syntactic relations had impact on the ease of error detection.

• Type: Conference
• Year: 2007
• Where: NAACL
• Authors: Yongmei Shi, Lina Zhou
• Keywords: Computational Linguistics, Error Detection, NAACL 2007, SR Errors, Syntactic Relations