The Association of Loudspeaker Manufacturers & Acoustics International presents MEASUREMENT OF HARMONIC DISTORTION AUDIBILITY USING A SIMPLIFIED PSYCHOACOUSTIC MODEL Steve Temme, Pascal Brunet, and Parastoo Qarabaqi of Listen, Inc. 2010 Listen, Inc.
Outline Motivation Our Perceptual THD Algorithm Algorithm Training & Experimental Results Conclusions
Motivation Conventional distortion measurements such as THD do not show reliable correlation to the ear s perception Our previous papers focused on Perceptual Rub & Buzz distortion (e.g. high order harmonic families) and has proved to be reliable at measuring perceived defects on the production line. We want a standard method for measuring the audibility of harmonic distortion in general both low and high orders.
PEAQ Standard ITU-R Recommendation BS.1387 (PEAQ) PEAQ= Perceptual Evaluation of Audio Quality Takes a reference signal (stimulus) and a test signal (response) as inputs Applies the effects of several psychoacoustic mechanisms to the signals Model Output Variables (MOVs) are generated that describe different properties of the reference and the test signals.
PTHD vs PEAQ Our PTHD Algorithm: Algorithm for the measurement of Perceptual Total Harmonic Distortion (PTHD) Based on the PEAQ standard Incorporates its psychoacoustic model MOV s selected for measurement of audible distortion in loudspeakers and headphones
Overall Concept DUT Stimulus Tone Response PTHD + 0 Human Listener - The goal is to replace the Golden Ear with a measurement algorithm.
Block Diagram of the PTHD Algorithm Stimulus & Response Spectra Calculation of the Model Output Variables Neural Network Perceptual Distortion Index
Some of the MOV S Total Noise Loudness: loudness of added distortion in the presence of masking Error Harmonic Structure: detects extended family of harmonics using Cepstral analysis Noise To Mask Ratio: ratio of the distortion level to the masking threshold Probability of Detection of the distortion by average listener Difference in Sharpness: detects addition of high frequency components
Neural Network Input layer M Scaled MOV(1) W x (1,M) Scaled MOV(1) W x (1,1) W x (N,1) Input layer 1 W y (M) Output layer Scaled MOV(N) B x (1) W x (N,1) W y (1) PDI B x (1) B y Non-linear Activation Function Neural Networks can approximate any arbitrary function
Acoustic Measurement Setup
Listening Tests Reference Scale Perceived Distortion Index (PDI) Description 1 Good- no perceived distortion. 2 Not as good- e.g. the signal is not as clear. 3 4 Distortion perceived- e.g. in the form of pitch and/or timbre change. Objectionable distortion- existence of multiple tones perceived. 5 Terrible distortion- e.g. Rub & Buzz.
Training of Algorithm Feedback DUT Stimulus Tone Response PTHD (Neural Net.) + Human Listener - The Neural Network is optimized to match its output to the reported PDI values by adjusting its parameters.
Experimental Results PDI= 1.5 (good) PDI= 2.25 (not so good) PDI= 4.5 (terrible) 80 THD&MOV vs Test Level 70 60 EHS [%] 50 40 THD [%] 30 20 NMR [db] 10 75 85 95 [dbspl]
Noise to Mask Ratio Distortion level to the masking threshold
Correlation with PDI THD PTHD PTHD clearly gives better correlation to perceived distortion level!
Conclusions An objective method is proposed for measuring the perceptual harmonic distortion of audio systems. The method uses a framework similar to the PEAQ standard. The measured signal qualities (MOVs) used are specifically selected to measure the harmonic distortion Subjective tests, so far, confirm the effectiveness of the proposed method. Future work: use real signals not test tones!