Noise in Speech-to-Text Voice: Analysis of Errors and Feasibility of Phonetic Similarity for Their Correction

Abstract

In Australian healthcare, failures in information flow cause over one-tenth of preventable adverse events and are tangible in clinical handover. Regardless of a good verbal handover, anything from two-thirds to all of this information is lost after 3–5 shifts if notes are taken by hand or not taken. Speech to text (SST) and information extraction (IE) have been proposed for taking the notes and filling in a handover form with extrapolated evaluations from related studies promising over 90 per cent correctness for both STT and IE. However, this cascading evokes a fruitful methodological challenge: the severe implications that errors may have in clinical decision-making call for superiority in STT; the correctness percentage measured in a peaceful laboratory is decreased to 77 by noise in clinical practise; and the STT errors multiply when cascaded with IE. We provide an analysis of STT errors and discuss the feasibility of phonetic similarity for their correction in this paper. Our data consists of one hundred simulated handover records in Australian English with STT recognising 73 per cent of the 7,277 words (1 h 8 min 5 s) correctly. In text relevant to the form, 836 unique error types are present. The most common errors include inserting and, in, are, arm, is, a, the, or am (5≤n≤94), deleting is (n = 17), and substituting and, obs, are, 2, he with in, also, to, or and she (7≤n≤11), respectively. Eighteen per cent of word substitutions sound exactly the same as the correct word and 26 per cent have a similarity percentage above 75. This encourages using phonetic similarity to improve STT. Show more