Bitter Pills to Swallow. ASR and TTS have Drug Problems

In an emerging application of speech technology, physicians speak prescription orders into wireless handheld devices that record the order using automatic speech recognition (ASR). Orders are recognized using ASR, then displayed on the screen for verification and, after checking, are relayed to the pharmacy for filling. Given the frequent opportunities for mistakes, this scenario provokes chills in many of us. Many physicians do not speak English as a native language, or they have a strongly accented variety of English. Some drugs are familiar enough that they should not present recognition/pronunciation difficulties (e.g. aspirin, codeine). Hundreds of drug names are easily confusable; other names range from opaque to ambiguous to 'no idea' with regard to pronunciation. The problems with the 'unknowable' drug names (the majority) are sometimes alleviated by drug manufacturers providing a 'pronunciation hint', in a dictionary-style phonetic transcription. This is not necessarily helpful to native or non-native speakers of English, or those unfamiliar with such renditions. More often, we are left to our own (wobbly) intuitions about stress placement, vowel length or 'hard' or 'soft' letter 'c', etc. Speaking the potentially infinite class of drug names correctly is an extra challenge for any TTS system. This paper presents data for drug names in the confusable, transparent, opaque, and unknowable categories. The clear and present dangers for patients receiving the wrong drugs and for everyone using such speech-automated systems are examined.