Assessing the efficacy of benchmarks for automatic speech accent recognition

Speech accents can possess valuable information about the speaker, and can be used in intelligent multimedia-based human-computer interfaces. The performance of algorithms for automatic classification of accents is often evaluated using audio datasets that include recording samples of different people, representing different accents. Here we describe a method that can detect bias in accent datasets, and apply the method to two accent identification datasets to reveal the existence of dataset bias, meaning that the datasets can be classified with accuracy higher than random even if the tested algorithm has no ability to analyze speech accent. We used the datasets by separating one second of silence from the beginning of each audio sample, such that the one-second sample did not contain voice, and therefore no information about the accent. An audio classification method was then applied to the datasets of silent audio samples, and provided classification accuracy significantly higher than random. These results indicate that the performance of accent classification algorithms measured using some accent classification benchmarks can be biased, and can be driven by differences in the background noise rather than the auditory features of the accents.