Evaluating the Performance of ASR Systems for TV Interactions in Several Domestic Noise Scenarios

Voice interaction with the television is becoming a reality on domestic environments. However, one of the factors that influences the correct operation of these systems is the background noise that obstructs the performance of the automatic speech recognition (ASR) component. In order to further understand this issue, the paper presents an analysis of the performance of three ASR systems (Bing Speech API, Google API, and Nuance ASR) in several domestic noise scenarios resembling the interaction with the TV on a domestic context. A group of 36 users was asked to utter sentences based on TV requests, where the sentences' corpus comprised typical phrases used when interacting with the TV. To better know the behavior, performance and robustness of each ASR to noise, the tests were carried out with three recording devices placed at different distances from the user. Google ASR proved to be the most robust to noise with a higher recognition precision, followed by Bing Speech and Nuance. The results obtained showed that ASR systems performance is globally quite robust but tends to deteriorate with domestic background noise. Future replications of the evaluation setup will allow the evaluation of ASR solutions in other scenarios.