approaches for testing and evaluating car audio … schäfer approaches for testing and evaluating...
TRANSCRIPT
Magnus Schäfer
HEAD acoustics GmbH
ETSI STQ Workshop on Multimedia Quality
in Virtual, Augmented and other Realities
Sophia Antipolis, 10.05.17
Approaches for Testing and Evaluating
Car Audio Systems
Approaches for Testing and Evaluating Car Audio SystemsMagnus Schäfer 2
For most car drivers and passengers today, the enjoyment of
music is an inherent part of their driving experience.
Continually increasing customer expectations regarding audio
quality
How to evaluate sound quality as perceived by the user?
Auditory testing
Instrumental evaluation
Introduction
Approaches for Testing and Evaluating Car Audio SystemsMagnus Schäfer 3
Auditory Test Design – Recordings
Auditory evaluation of car audio systems
Concept is based on real music playback and DF-equalized
binaural recordings of head and torso simulator (HATS)
Measurements were conducted using common cars and their
original audio systems
Real (professional) music tracks as test stimuli: rock, pop, classic,
jazz, r&b, … Playback via CD, 7 tracks per DUT
Approaches for Testing and Evaluating Car Audio SystemsMagnus Schäfer 4
Auditory Test Design – Procedure
Several auditory test methods and scales of different standards are
available for testing audio quality:
ACR, DCR, CCR, MUSHRA, …
Focus on large amount of test conditions
ACR test with 5-point scale acc. to ITU-T P.800
Loudness normalization to 23 sone (N5 acc. to ISO 532-1)
Hearing-adequate playback of binaural recordings
Anchors, arbitrary level offsets and other files included
161 samples in overall
Usually tests are conducted in a listening laboratory
Psychologic aspect of the listening test so far not considered:
Do test subjects behave differently in a real car?
Is the quality perception the same inside a real car?
Approaches for Testing and Evaluating Car Audio SystemsMagnus Schäfer 5
Testing Environments – Laboratory
Everything is under control (light, seat, air condition, …)
very „clean“ environment
Idle noise < 20 dB(A)
Several terminals can be utilized simultaneously
But is it too unrealistic / artificial for car audio evaluation?
Approaches for Testing and Evaluating Car Audio SystemsMagnus Schäfer 6
Testing Environments – Driving simulator
More „lively“ environment: street is projected on a screen in front of
the car (interaction possible but not part of the current evaluation)
Idle noise < 20 dB(A)
Only one test person at once
Not always accessible for subjective testing
More realistic for auditory evaluation?
Approaches for Testing and Evaluating Car Audio SystemsMagnus Schäfer 7
Results of Auditory Test
Comparison of testing environments (details: Reimes et al. DAGA 2017)
Results highly correlate (RMSEStimuli = 0.236, RMSEDUT = 0.103)!
Level of significance of differences (ANOVA / t-Test) …
DUT 𝑝 = 0.84
Stimuli 𝑝 = 0.48
Votes 𝑝 = 0.046∗ (Cohen’s d: 𝑑 = 0.06 ≪ 0.2 not relevant)
No systematic under- or overestimation of perceived quality
Approaches for Testing and Evaluating Car Audio SystemsMagnus Schäfer 8
Record audio signal at driver position
Comparison between input signal and recording
Focus here for illustration: Spatial properties
Approach for Instrumental Evaluation
Evaluation procedure
Audio recording
Binaural hearing model
Input and output
Analysis of differences
Mapping to perceptual scale
Approaches for Testing and Evaluating Car Audio SystemsMagnus Schäfer 9
Coincidence-based model
Cross-correlation with inhibition [Lindemann86]
Frequency and lateralization weighting [SchäferBahramVary12]
Output: Correlogram
Additional derivative outputs possible
Binaural Hearing Model – Structure
Weighting and Averaging
Inhibited cross-correlation
Monaural ear modelMonaural ear model
Approaches for Testing and Evaluating Car Audio SystemsMagnus Schäfer 10
Example: Correlograms for two car audio systems
Reference input signal
Classical music
Binaural Hearing Model – Output
Good performance Bad performance
Approaches for Testing and Evaluating Car Audio SystemsMagnus Schäfer 11
Difference (and absolute values thereof) between correlograms of
recording and reference
Mean
Median
Percentiles
5th, 90th and 95th
Combination of several metrics
Random Forest Regression
Metrics and Regression
Approaches for Testing and Evaluating Car Audio SystemsMagnus Schäfer 12
Entire listening test results used for training
Trained combination of all metrics can reproduce auditory results
Regression Results
memorize?
Approaches for Testing and Evaluating Car Audio SystemsMagnus Schäfer 13
Listening test results split 50-50 between training and evaluation
Better than individual metrics – but far from satisfactory
Regression Results
Approaches for Testing and Evaluating Car Audio SystemsMagnus Schäfer 14
Conclusions
Setup of auditory testing for car audio systems based on binaural
recordings and hearing-adequate playback
Large auditory evaluation was conducted with 161 stimuli in two
test environments with 45 participants
Results between listening environments provide high correlation
No systematic differences can be observed in this study
Further auditory tests can be conducted in listening lab!
Spatial analysis of recordings from different car audio systems
Binaural Hearing Model
Neither individual metrics nor trained regression provide
satisfactory overall quality estimation
Using spatial properties is not enough
Magnus Schäfer
Telecom Research & Standardization
www.head-acoustics.de © Copyright HEAD acoustics GmbH