Measurement of Car Audio Sound Quality using Audio Content Features

Ranking the sound quality of car audio systems in an industrial environment is an important activity used to bench-mark a product among its competitors and to qualify a possible added-value of the product for marketing purposes, and can also help in the development process of an ongoing project. But even at static conditions, it is a complicated, expensive and time consuming task which is typically carried out by means of listening tests, objective methods like impulse response measurements or even combination of both using in-car binaural recordings with dummy heads. Such methods have to face a large number of practical issues and are still arduous and time consuming. The work presented in this thesis corresponds to the initial developments of a predictive perceptual model intended to output sound quality rankings of car audio systems from the audio content analysis of in-car binaural recordings. The model is calibrated using machine learning techniques on a data base built from the measured scores resulting from listening tests using binaural signals and the audio content features extracted from the same audio signals. The main objective is to provide a tool for fast product benchmarking in an industrial context. The proposed methodology combines techniques belonging to four different disciplines, namely binaural technique, perceptual audio evaluation, audio content processing and data mining. Some of the audio content processing and machine-learning techniques applied are typically used in Music Information Retrieval research. Other techniques have been borrowed or inspired from state-of-the-art methods in perceptual modeling, including ITU-R standards, binaural auditory modeling or psychoacoustics. Binaural recording and reproduction techniques combined with standard perceptual evaluation methods were applied in the design of a new experimental paradigm that seems to work well to carry out listening tests in this context. A set of sensory attributes was chosen in order to guide both the listening tests and the design of audio content processing blocks to extract perceptually motivated audio descriptors. The resulting dataset was trained using different learning schemes some of which yielded good cross-validation results. Among them, an M5Rules algorithm was chosen, showing quite good predictions for five of the seven tested cars when validated using previously unseen data. Although more subjective data is still to be collected for a proper model training and validation, the preliminary results obtained can be considered as promising enough to continue the work in the direction of the presented approach. In this way, some guidelines for next steps and future research have been identified and proposed.