ABSTRACT
Most automated speech scoring systems, including SpeechRater, function by first extracting multiple features from the audio recording of the response and the transcription generated by an automatic speech recognition system. These features provide a numeric representation of different linguistic properties of the response and are then combined into a single proficiency score using rules, a mathematical formula, or a combination of both. In SpeechRater, different features are designed to correspond to different aspects of linguistic performance, such as pronunciation, vocabulary sophistication, and overall fluency, in addition to others. With multiple construct-relevant features available, there are two important decisions involved in the SpeechRater scoring model design process: what features should be used in the final model and what weight should be assigned to each feature. Various operational and research automated constructed response scoring systems have explored the use of methods other than linear regression.
