Sounds great

Customers’ expectations regarding audio quality are continually increasing – something car manufacturers must pay attention to. However, it’s not an easy task to evaluate the quality of delivered audio systems.

The challenge

The main validation process normally involves a series of time-consuming subjective audio evaluations. This means that an exorbitant amount of hours are spent checking and controlling the quality of installed systems, and it is difficult to link the results of subjective evaluations to possible solutions. To help car manufacturers figure out whether the delivered system meets their expectations and, if not, to see which perceptual attributes need to be improved, a new prototype tool has been developed based on a limited set of sound samples.

Non-intrusive method

This project developed an audio-quality algorithm for multi-channel audio systems in vehicles. It’s a non-intrusive method, which means that the process doesn’t use a reference sound sample, which can’t be obtained for audio systems in cars. The algorithm is based on spatial metrics, distortion metrics, and sound quality metrics.

To develop this algorithm, one classical music sample and two pop-music samples were recorded inside vehicles under test conditions, and expert listeners identified six perceptual attributes. The recorded sound samples were then played in a listening room to twenty expert subjects who evaluated the overall quality, as well as the six perceptual attributes.

The tests revealed the overall preference and the perceptual attributes of the recorded sound samples. Subjective results were then predicted by a linear regression involving the principal components of the objective metrics. Furthermore, a new distortion metric was developed to detect and quantify any non-linear distortion generated by the loudspeaker systems, for example, rattling of the panels.

Teachable system

The result is prototype software that calculates the relevant objective metrics and predicts the perceived attributes. This is a teachable system, so new subjective data can be added in the model to further improve prediction accuracy. Whereas other algorithms require a reference sample (with audio quality close to perfect), this evaluation technique is advantageous for car audio, where it is impossible to know what the perfect system is.

Once there is enough subjective data, the listening test results can be applied to other car audio systems, without performing further listening experiments. This approach allows for more effective evaluation of multi-channel audio in vehicles, without performing  time-consuming subjective evaluations.