Evaluation of the Perceived Audio Quality of Hearing Aids

Transcription

1 DSP Application Day 2006 Milano - September 18, 2006 DICo Dipartimento di Informatica e Comunicazione Università degli Studi di Milano Evaluation of the Perceived Audio Quality of Hearing Aids Antonio Mancuso, Lorenzo Picinali, Giancarlo Vercellesi LIM/DICo Università degli Studi di Milano knmbm@tin.it, lpicinali@dmu.ac.uk, vercellesi@dico.unimi.it Abstract: The higher cultural and social level of people affected by hypoacusis, leads to a growth of the request of high quality hearing aid systems: to achieve this quality enhancement, the evolution of technologies in the hearing aid and audiometric fields is moving towards the digital world. Even if digital audio prosthesis are not new, the algorithms for the enhancement of speech quality and intelligibility are always being improved. We propose a methodology for the evaluation of the perceived quality of hearing aids: the simulation of a real two-dimensional soundscape is done using a quadraphonic loudspeaker array, then a speech signal (extracted from the Italian Audiometric Vocal Audio Archive) is reproduced from a single loudspeaker placed in different positions around the virtual soundscape, and everything is recorded through a Dummy Head microphone system with hearing aids placed on the two ears. The analysis of the signal recorded in these simulations is carried out through a complex comparison of various parameters extracted from the signals in output from the Dummy Head microphone and from the hearing aids. 1. Introduction The evolution of technologies of the hearing aid and for the audiometric fields is going toward the digital. Furthermore, there is a growing request of better hearing aids systems, which should satisfy the increasing cultural and social level of people affected by hypacusia. In this paper we propose a methodology to evaluate the perceived quality hearing aids. We simulated the real world using a quadraphonic system to reproduce the sound, the Italian audiometric vocal audio archive to reproduce a monophonic speech signal and the dummy head as input system. We got audio files from dummy head and hearing aids. Then we analyzez their quality extracting the standard audio parameters and comparing them with the PESQ (Perceptual Evaluation of Speech Quality) evaluation algorithm. Finally, we performed subjective tests on normal and hypacusic people and we compared them with the objective evaluation in output from various algorithms.

2 2 2. The Digital Hearing Aids Digital hearing aid is a hearing aid device that receives sound and digitizes it (breaks sound waves up into very small, discrete units) prior to an amplification. A traditional analog hearing aid simply makes the sound wave larger to amplify sounds. Like their analog counterparts, digital hearing aids are available in a range of prices, and with a corresponding range of capabilities. The most sophisticated digital hearing aids have built in intelligence that allows them to discern between soft, but desirable sounds, and louder, but unwanted noise. Such devices can amplify the former, while neutralizing the latter for better performance in a variety of environments. A digital hearing aid can be programmed to adjust itself to the current environment millions of times each second. Digital technology also makes it possible for technicians to create customized programs that address each individual's specific hearing difficulties. An analog hearing aid employs miniature devices, such as microphones, amplifiers, and resistors to alter an acoustic signal (a sound) electronically. The altered signal is then converted back to a sound wave and played into the ear canal. A digital hearing aid also uses a microphone, and a microprocessor that performs most of the device's functions digitally. Because the signal is represented as a series of numbers, it can be quickly and accurately altered [10]. Speech clarity and sound quality are the most important goals that digital hearing aids, which should reach them by conveniently digital signal processing. Generally, the speech clarity is provided by: A conveniently enhancement of directionality providing appropriate gain for varying levels of speech and boosting SNR (signal-to-noise) through directional processing. An adaptive directionality with gradual directional switching An algorithm which implement the feedback cancellation in both microphones. The sound quality may be provided by: A seventeen-band compression and noise reduction with frequency warping A noise reduction via spectral subtraction

3 3 Furthermore, the fitting accuracy is provided by the data-logging and analysis of hearing instrument use. And finally, it is important to improve the comfort wearing via open hearing instrument by conveniently technologies as Stabilizer DFS. 3. The Perceived Quality Evaluation The state-of-the art of methods and algorithms used to evaluate objective and subjective perceived quality are described in the following lines. Objective quality: a short overview of PEAQ and PESQ architectures are presented; the implementation illustrated in [1] and [8] has been used to compute the objective quality indexes. Subjective quality: recommendations provided by [2], [3] and [5] have been analyzed; a short description of the subjective tests method is presented together with the description of the software and criteria used to select listeners. Finally, [14] discuss some procedures to evaluate the perceived audio quality of hearing aids. 3.1 The Objective Quality Evaluation The recommendation [1] and [8] defines a method to compare a reference signal and a signal under test. The output are called MOS (Mean Opinion Score) for PESQ [8], ODG (Objective Difference Grade) for PEAQ [1]. These are estimation of the perceived difference, and they vary between -0.5 and 4.5 for PESQ and between 0 to 5 for PEAQ. For both systems, the estimation is based on: A sophisticated ear model comprising several intermediate steps. Two different ear models are provided: one based on FFT and one based on filters bank The calculation of a conveniently set of psychoacoustics variables. E.g. in PEAQ system they are called MOVs (Model Output Variables) A mapping from a set of psychoacoustics variables to a single value (computed by a neural network) representing the perceived difference. They are believed to be representative of the audio quality (no difference = good quality, high difference = bad quality) In the case of stereo signals, all the computation stages are performed independently for the left and right channel, and then a mean values is calculated, except where otherwise indicated.

4 4 We have selected a set of reference CD-quality signals from the Italian vocal audiometric [9]. As noise signal, we have used pink noise, cocktail party and traffic. Misalignments may be introduced by the Dummy Head acquisition: e.g. null samples at the beginning or at the end of the test file. These misalignments affect the objective evaluation. We therefore manipulated the test files in order to have a perfect alignment. 3.2 The Subjective Quality Evaluation ITU-R provides many methods to perform subjective listening tests. The most relevant: For general audio quality (PEAQ system) are described in [2], [3] and [5]. For voice quality (PESQ system) are described in [6] and [7]. [2] describes a method for the subjective assessment of small impairments using the double blind triple-stimulus with hid-den reference. It employs a 5-point impairment scale with anchor: 1 Very Annoying, 2 Annoying, 3 Slightly An-noying, 4 Perceptible but not Annoying, 5 Imperceptible. [3] defines a more general guide for evaluation of subjective perceived quality. It is based on [2] and generally is indicated in detection of large impairments. [5] describes the so called MUSHRA test. It is aimed to handle intermediate audio quality using the Multi Stimulus Hidden Reference and Anchor method. It employs a 5-point impairment scale with anchor: 1 Bad, 2 Poor, 3 Fair, 4 Good, 5 Excellent. [6] and [7] describe recommended procedures for conversational and listening-only methods of subjective evaluation. Laboratory conversation tests are intended as far as possible to reproduce, in the laboratory situation, the actual service conditions experienced by customers. Listening-opinion tests are not expected to reach the same standard of realism as conversation tests, and the restrictions are therefore less severe in some respects. Nevertheless, the artificiality that has to be accepted brings with it a necessity for strict control of many parameters which in, conversation tests, are allowed to find their own equilibrium. The recommended test method for listeningonly tests is the Absolute Category Rating, the Quantal-Response Detectability Method, the Degradation Category Rating, the Comparison Category Rating and the Threshold Method.

5 5 4. Hearing Aids Test Platform To carry out all the objective and subjective testing experiments, it is necessary to establish a hardware and software platform for the extraction of the signals that have to be evaluated. These signals can be divided in two different kinds: the first ones have to come out from the prosthesis, so they have to be picked up at the end of the little pipe that carries the sound from the prosthesis s transmitter to the ear canal. The second ones are the signals that represent what the listener can hear in normal conditions, without the hearing aid. To obtain this second kind of signals, one first option could be to put two miniaturized microphones at the beginning of the two ear canals of an human listener: of course this option would create various problems, due to the impossibility for the listener to stay exactly in the same position for a certain number of hours necessaries for the performing of the experiments. The second option, the one that we adopted for the development of the experiments, is to use a dummy head. 4.1 The Dummy Head The dummy head is a mannequin head, made in latex filled up by polystyrene, that has the same dimensions of a human head (nearly 54 cm of horizontal circumference at the level of the two ears, 59 cm of vertical circumference at the level of the two ears and 17 cm of distance between the two ears). At the entrance of the two ear canals of the dummy head, we placed two miniaturized condenser microphones (one for each ear). On the market, is it possible to find a certain number of dummy heads (from Neumann, Kemar, Cortex, and various other brands), but none of these were valuated as sufficiently precise for the realization of the goal of our experiments. At the end, we decided to build a customized dummy head, using two precise casts of a real pinna and choosing the most suitable miniaturized microphone for the recording of the signals. In Figure 1, it is possible to observe the dummy head used for the experiments, while in Figure 2 it is showed a detail of the pinna with the miniaturized microphone capsule.

6 6 Figure 1: the custom dummy head used for the experiments Figure 2: a detail of the pinna with the microphone capsule placed at the beginning of the ear canal of the right ear

7 7 4.2 The HW/SW Platform Starting with the software platform, for the carrying out of the testing there have been used the following systems: Noah 3, with the module Aventa 2.05 and Audiometer: this is the software used for the calibration of both the hearing aids and for the extraction and calculation of the hear sensibility of some representative subjects (samples ). Digidesign Pro Tools LE 6.7 and 7.0: this is the software used for the reproduction of the two dimensional soundscape (quadraphonic noise and monophonic speech) and for the recording of the signals coming out from dummy head and from hearing aids. Audio Test, Csound, Sound Hack 0.093, Audio Sculpt PPC 2.1 and Logic Audio Pro 7: these software have been used for the recording, the generation and the calibration of the audio signals during the experiments. MAX-MSP-Jitter: this is the object oriented programming language used for the developing of the subjective testing platforms for the testing of the signals in output from the experiments. Then, here follows a brief list of the hardware equipment: Sennheiser ke-4: these are the first microphone capsules used for the dummy head. They have been substituted after a few months because of the noise that these generate; the main problem was the custom power supplies, which could not allow a good SNR (Signal to Noise Ratio). DPA 4060bm: these are the miniaturized condenser capsules used for the dummy head (after the Swnnheiser ke-4). Briefly, the specifications of these capsules are: Frequency response: Hz +-1 db Power supply: 48v Max SPL: 130 db SNR: 110 db Sennheiser MK 2-ew Gold: these are the microphones used to collect the output signal from the two hearing aids.

8 8 Digidesign 002 Rack: this is the audio interface (microphone preamplifiers and AD-DA converters) used for the experiments. Briefly, the specification of the interface are: Sample Rate: 96 khz Bit Resolution: 24 bits Microphone preamplifiers: 4 Analogue and digital input: 18 Analogue and digital output: 18 Bus used for the communication with the computer: firewire 400 Aurical Plus: multifunctional audiometer system used (in couple with the Audiometer module inside the Noah 3 software) for the calibration of the hearing aids and for the calculation of the hearing curves of representative subjects. Phonic PAA2: digital phonometer and 31 bands spectrum analyzer, for the analysis and the calibration of the equipment used for the experiments. Yamaha MSP3: active two way loudspeakers used for the reproduction of the two dimensional soundscape. PC PIV with Windows 2000: computer used for the calibration software. Apple Powerbook G4 1 Ghz: computer used for the reproduction and the recording of the audio signals. 4.3 The Test Performed It would be impossible to describe shortly in these pages all the tests performed for this research project in the last month; we could just have a brief overview of some of the most important tests carried out: Extraction of the Polar Pattern: in this stage of the experiment, it has been extracted the polar pattern of both the hearing aids with the Active Directionality function turned on. With the reproduction of a sinus-logarithmic sweep signal in various positions around the dummy head with the prosthesis placed on both the ear, it was possible to reconstruct the polar pattern of both the hearing aids for various frequencies. Comparing it with the one extracted form the dummy head, it was possible to notice that the hearing aids have a particular directional enhancement for the sound signals coming from the front of the listener, with a really strict hyper-cardioid polar pattern.

9 9 Testing of the Noise Reduction algorithm: in this stage of the experiment, it has been tested the efficiency of the noise reduction algorithm (called Noise Tracker ) of both the hearing aids. Recreating a real world situation of a twodimensional noise signal that masks a monophonic speech, it was possible to compare the signals in output from both the dummy head and the two prosthesis and to notice that, while in the dummy head output the speech signal was not completely noticeable (because it was masked by the quadraphonic noise), in the hearing aids output it was clearly possible to see the reduction of the masking noise (even until db of reduction in less than 3 seconds) and the enhancement of the speech, that obviously was clearly audible even in critical situations. Comparison of the performance of three different kinds of hearing aids: in this stage of the experiment, it has been performed a comparison between three different kinds of audio prosthesis (two of them were based on a digital engine, and one was based on an analogue engine) with all the sound processing algorithm deactivated (to test just the response of the flat prosthesis). Extracting the parameters described in the previous stages, and some more parameters about the frequency response of the AD-DA conversion and of the internal processing of the hearing aids, it was possible to compare the data and to make a list of the better working heading aids (for each algorithm and function tested) for different masking situations.

10 10 5. Analysis of Results We have performed many tests on various hearing aids available from the market. In particular, we have selected the GNResound products which provide many algorithms to enhance the speech clarity, the sound quality and the fitting accuracy. More precisely we have selected Metrix and Canta7 instruments. Following the protocol defined in the previous section, we kept under stress the acoustic prosthesis defining several tests, some of which have been briefly defined previously. In Table 1 the experiments are shortly listed: Test Type Test Description Technical Analysis Objective Evaluation Adaptive Directionality Ability of hearing aids to change in realtime the polar pattern for a directional Polar Diagram PESQ Index Noise Tracker Flat frontal focusing Ability of hearing aids to detect the voice, reducing the background masking noise Hearing aids behavior with all processing algorithms deactivated Table 1: List of the performed test on hearing aids Waveform, Sonogram and Numerical Analysis Waveform and Sonogram PESQ Index PESQ Index For each test, we provide the analysis of results with the related sonograms, waveforms, polar diagrams and PESQ analysis.

11 Adaptive Directionality with sweep Figure 3: polar diagram of Adaptive Directionality test for the 562 Hz frequency: here it is clearly noticeable the strong directionality of the hearing aid, oriented towards the frontal side of the listener s soundfield. In this case, the response considered is just for the frequency of 562 Hz.

12 Figure 4: polar diagram of Adaptive Directionality test for all frequencies: also in this polar diagram it is clearly noticeable the strong directionality of the hearing aid, oriented towards the frontal side of the listener s soundfield. In this case, the response considered is for all the frequencies from 200 to Hz.

13 Noise Tracker with Frontal Voice and Cocktail Party Figure 5: waveforms of Noise Tracker test with cocktail party: the fist oscillogram represent the signal in output from the dummy head, while the second one represent the signal in output from the Metrix hearing aid. It is clearly noticeable the reduction of the background noise brought by the Noise Tracker algorithm after just a few second from the beginning of the signal; it is also important to notice that the speech has not been involved by the reduction, and that his average level remains the same.

14 14 Figure 6: sonograms of Noise Tracker with cocktail party: the fist sonogram represent the signal in output from the dummy head, while the second one represent the signal in output from the Metrix hearing aid. Also in these pictures it is possible to notice the reduction of the background noise after a few seconds from the beginning in the hearing aid output.

15 15 Noise Tracker 3,00 2,50 PESQ Index 2,00 1,50 1,00 DUMMY METRIX 0,50 0,00 COCKTAIL PARTY TRAFFICO RUMORE Tipologia di Rumore Mascherante Figure 7: objective analysis of noise tracker test by PESQ with different noises. It is possible to notice that Metrix has a better quality than dummy head in all the tested situations.

16 Test Comparison Flat with Cocktail Party Figure 8: waveforms of flat comparison test with cocktail party: the first one is the output of the dummy head, the second one is the output of the analogue prosthesis and the third one is the output of the GNResound-Metrix prosthesis. It is possible to notice the differences, between the three oscillograms, talking about the sensibility to the noise and the sensibility to the speech: for example, the third signal (the one in output from the Metrix, the most advanced prosthesis tested here) seems to show a reduction of the background noise even if the algorithms for the noise reduction are turned off.

17 Figure 9: sonograms of flat test with cocktail party: the first one is the output of the dummy head, the second one is the output of the analogue prosthesis and the third one is the output of the GNResound-Metrix prosthesis. Also here it is possible to notice the different response of the three systems tested. 17

18 18 FLAT 3 2,5 PESQ Index 2 1,5 1 DUMMY ANALOGICA METRIX 0,5 0 COCKTAIL PARTY RUMORE TRAFFICO Tipologia di Rumore Mascherante Figure 10: objective analysis of flat test by PESQ with different noises. In all the situations (different masking noises), the Metrix has a better quality than the Dummy Head. It is interesting to note that Analogue one works better than Metrix and Dummy Head. We think that these results are caused by the analog to digital and digital to analog conversion performed by digital hearing aids (all these conversions could bring some latency, which can be badly valuated by the PESQ algorithm). 6. Conclusion and Future Works In this paper we have proposed a methodology to evaluate the perceived quality hearing aids. By using a quadraphonic system to reproduce the sound, the Italian audiometric vocal audio archive as test files and the dummy head as input system, we have tried to simulate real world situations. We got audio files from dummy head and hearing aids, then we analyzed their quality extracting the standard audio parameters and comparing them with the PESQ (Perceptual Evaluation of Speech Quality) evaluation algorithm. Finally, we performed subjective tests on normal and hypacusic people and we compare them with the objective ones. It could be difficult to say now what our results have shown: the attempt (the final goal) of this research project is to establish an objective and subjective testing methodology for the evaluation of the perceived quality of the audio prosthesis, so we can say that in these first steps we performed different kind of test and we decide which one could be the most suitable for the definition and evaluation of the perceived quality. The results (the evaluations) of these tests have shown the

19 19 difference between a listening situation with and without the hearing aids, and have underlined certain properties of the hearing aid signal processing and enhancements brought by the audio prosthesis for the improvement of the perceived quality. Even if we are still far away from the formalization of a definitive testing methodology, it is possible to say that after this first year of research we achieved important results, and that in the following steps of the research project we will go on in the same direction, taking in consideration also parallel paths like the testing of noise-cancelling headphones, the developing of testing platforms for the implementation of various methodologies for the quality evaluation of sound signals, the auditory scene analysis (ASA) and the intelligibility of music. Acknowledgments This project has been developed in the context of agreement between GNResound and University of Milan. The authors say thanks to GNResound for providing the audiologist equipments and facilities and for allowing us to present these results. A special thanks to dott. Fabio Gomiero and dott. Gianluca Vivarelli of GNResound. Prof. Goffredo Haus (scientific director of LIM and director of DICO) and Davide Mauro of LIM. Prof. Silvano Prosser of University of Ferrara. References [1] ITU Recommendation, ITU-R BS , Method for Objective Measurements of Perceived Audio Quality, 2001 [2] ITU Recommendation, ITU-R BS , Methods For The Subjective Assessment of Small Impairments in Audio Systems Including Multichannel Sound System [3] ITU Recommendation, ITU-R BS , General Methods For The Subjective Assessment of Sound Quality [4] P. Kabal, An Examination and Interpretation of ITU-R BS.1387: Perceptual Evaluation of Audio Quality, McGill University, 2002 [5] ITU Recomendation, ITU-R BS.1534, Method for the subjective assessment of intermediate quality level of coding systems [6] ITU Telecommunication, ITU-T P.800, Methods for Subjective determination of transmission quality [7] ITU Telecommunication, ITU-T P.830, Subjective performance assessment of telephone-band and wide-band digital codecs [8] ITU Telecommunication, ITU-T P.862, Perceptual evaluation of speech quality (PESQ): An objective method for end-to-end speech quality assessment of narrow-band telephone networks and speech codecs [9] F. Cutugno, S. Prosser, M. Turrini, Audiometria Vocale, GNResound [10] H. Luo and H. Arndt, Digital Signal Processing Technology and Applications in Hearing Aids, Proc. IEEE 6th Int. Conf. on Signal Processing, vol. 2, pp (2002) [11] A. Markides, Binaural Hearing Aids, Academic Press, London, (1977) [12] Brian C. J. Moore, An Introduction to the Psychology of Hearing, Academic Press, London, UK (2003) [13] W. A. Yost, Fundamentals of Hearing: An Introduction, Academic Press, San Diego, California, USA (2000) [14] S. Lydia, D. Harvey, Y. Ingrid, W. David Procedure for Selecting the Saturation Sound Pressure Level of Hearing Aids: Experimental Validation, Ear & Hearing. 19(4): , August 1998.