Contribution
The future of speech quality evaluation in the presence of AI based implementations
* Presenting author
Abstract:
The evaluation of speech quality is a multidimensional task. Evaluation criteria have to take into account the various conversational situations: listening, talking and interactive conversation, they have to consider the acoustical situation where the talker and listener are in. This concerns background noise and acoustical room conditions. In consequence different types of test signals and different types of test scenarios were developed. The goal is always to define test procedures leading to meaningful, reliable, reproducible results.Developers of algorithms and terminals always try to improve the system performance by implementing the latest, state of the art technology to improve speech quality in all conversational situations. This stimulates the implementation of techniques automatically adapting to different acoustical situations. While until recently such techniques were just adaptive new implementations use self-learning AI based solutions.This may change testing strategies significantly. Self-learning algorithms will behave differently over time and ideally, they will improve their performance for their lifetime. However, there is the possibility that learning strategies may not enhance the users perception in all situations. This contribution discusses new test strategies, the question, whether the concept of minimum performance requirements may still hold for such systems and which potential alternatives could be seen.