Richard Baker University of Manchester Digital Signal Processing, compression, linear and nonlinear: terminology, measurement and issues. Richard Baker University of Manchester

Outline A few common misconceptions What is signal processing? Advantages of going digital Analogue to digital conversion Compression – why and how? Measurement issues

Common Misconceptions “Only digital hearing aids are signal processing aids” “Digital is better than Analogue” “Wide dynamic range compression (WDRC) = digital” “Nonlinear = digital” “Programmable hearing aids are the same as DSP hearing aids” “Digital hearing aids cut out background noise”

What is signal processing? Signal processing is exactly what it says, it may be: Amplifying Filtering Peak-clipping Compression: output limiting, WDRC, etc Frequency shifting … etc.

What is a digital hearing aid? A digital hearing aid simply converts the signal to a numerical form before processing it It’s the signal processing algorithm that is important

What is compression? Compression: the range of input sound intensities is “squashed” into a smaller range of output intensities e.g. a range of input intensities from 0 to 100 dB SPL may be compressed into an output range of 50 to 100 dB SPL The output “dynamic range” is reduced compared to that of the input

Why do we need compression? Sensorineural hearing loss most often results from damage to outer hair cells in the cochlear This results in: Loss of sensitivity at low sound intensities Abnormally rapid growth of loudness (recruitment) Loss of frequency selectivity (Hearing aids can’t do much about this one at the moment)

Loudness Growth Typically, sensorineural loss results in recruitment: Low intensity sounds are inaudible Moderate intensity sounds are heard as very quiet High intensity sounds are perceived as similar in loudness to that normal hearing listener Implications for hearing aids High gain for low intensity input Low gain for high intensity input i.e. reduced dynamic range at output compared to input

Compression Normal Impaired Non-linear Intense Moderate Weak Dillon (2001)

Hearing aid goals Audibility - be able to hear important sounds e.g. speech Comfort - sounds comfortably loud Safety - sounds prevented from being too loud Intelligibility - maximise the intelligibility of speech sounds Quality - maximise the perceived quality of the sounds (e.g. little distortion) Consistency - same performance regardless of listing conditions ... The same aims apply to both linear and nonlinear aids

Linear versus nonlinear Linear - gain is constant irrespective of input level (if we ignore very high levels) Nonlinear - gain changes as input level changes (may be compression or expansion) Remember, when talking in dB terms: Output level = Input level + gain

Linear hearing aids Amplify all sounds by the same amount Problem – louder sounds become too loud to be comfortable Solution – use some type of limiting to prevent this e.g. clip the peaks off the waveform when it goes too loud - peak clipping – causes distortion

Peak clipping

The need for compression The problem with linear aids – the same gain is applied to all levels of input signal we need high gain for low input levels, and low gain for high input levels - compression we need some way of automatically turning down the gain of the hearing aid as the input intensity increases an automatic gain control or AGC

Automatic gain control (AGC) AGC parameters Attack-time – The time taken for the AGC to respond to an increase in input level Release time – the time taken for the AGC to increase the gain again when the input level decreases Knee-point – below a certain signal intensity the amplifier behaves linearly, above this intensity the compression operates Compression ratio – above knee-point, output with an increase in input is typically less than 1 dB per dB change in input

Automatic gain control

I/O functions, output spectra & transfer functions etc. I/O functions - output vs input at one frequency Output spectra - output across frequency at one input level input/gain function - gain vs input Transfer function - output/input (i.e. gain) across frequency All ways of plotting different aspects of hearing aid function

Input-output function

Output spectra

Types of compression The main compression strategies fall into two categories: Compression limiting – high knee-point, high compression ratio (e.g. 10:1) – limits MPO WDRC – wide dynamic range compression, low knee-point, low compression ratio (e.g. 2:1) – aims to restore loudness perception in moderate loss AVC - automatic volume control - slow acting compression designed to adjust overall gain when moving from quiet to noisy environment.

Output limiting

WDRC

Therefore need to test at different levels: 50 dB SPL input - quite speech level 65 dB SPL input - moderate speech level 80 dB SPL input - loud speech level

Multi-channel processing Why multi-channel? different hearing losses at different frequencies different compression strategies required for different frequency ranges theoretical reasons for differing frequency response … … e.t.c.

From Killion et al, 1990

Test signals Pure-tone - single frequency component Swept-tone - pure-tone swept up or down in frequency Speech-weighted pure-tone sweep - swept-tone following the spectral shape of an average speech signal White-noise - noise signal containing equal energy at all frequencies Pink-noise - noise with energy decreasing with increasing frequency Speech-shaped noise - noise with spectral shape of an average speech signal Modulated Speech shaped noise - spectral AND temporal shape similar to that of speech

Test signals Test signals can be either: Continuous - long(ish) duration with approximately constant amplitude Fluctuating - varying up and down in amplitude (usually designed to mimic temporal fluctuations in natural speech) Least natural: continuous pure-tone Most natural: fluctuating speech shaped noise

Which signal to use? With a linear aid pure-tone test signals should produce the same results as noise signals With non-linear aids, the aid can respond very differently to different signals

Which signal to use? e.g. in some situations, pure-tones may produce an artificially high measurement of low frequency gain - “blooming” Suppose a compressor follows a high-pass filter A tone is swept upwards in frequency through the cut-off region of the filter into the pass-band As the tone is in the cut-off region the input to the AGC is low - thus the gain is high In the pass-band the input to the AGC is high so the gain is low Result: Using a swept tone it appears that the low-pass filter isn’t working – use a broad-band signal!

blooming! So, use a broad-band signal!

Which signal to use? e.g. swept-tone versus noise Pure-tone - single frequency component therefore level well defined White-noise - many frequency components - measured level is sum of frequency components therefore level at one particular frequency is lower Overall level with noise signal also depends on analysis bandwidth

Implications of different signals Output display for broadband signals is lower than tones - use gain display! Output display depends on analysis bandwidth For multichannel aids swept tone gives higher level signal through each band than broadband noise At high levels tone may result in saturation whereas noise doesn’t Nonlinear aids may have different gain for tones & noise even though they are nominally the same overall level

“extras” As well as different signal processing strategies modern hearing aids are available with many “extras” designed to improve their performance These also have implications for how the aids are tested and the signals used…

“extras” Noise suppression/cancellation Algorithms attempt to “detect presence of speech” and turn down the gain if no speech is present Note Need to use realistic speech like signal to perform measurements – continuous noise will be suppressed, so need to have speech-shaped noise with fluctuating envelope (is such a signal available?) Turn the noise reduction feature off

“extras” Multi-program/memory aids Can allow 2 or more different processing algorithms to be used E.g. a second setting with extra gain for bouts of OME Note Need to know what each of the memories are supposed to do in order to test aid

“extras” Directional/Multi-Microphone technology Aims to improve signal-noise ratio by “picking out” sounds from the front, and reducing those from other direction Note Need to be careful how aid is positioned in a test box to get accurate measurements Turn the directional microphone off!

“extras” Feedback management/cancellation Notch-filters or complex feedback cancellation algorithms have been developed that can reduce feedback and allow 10-20dB extra gain. This can allow additional gain, use of vents where they are normally not possible etc. Note: awareness of notch-filters is necessary & the feed-back suppression needs to be turned off for measurement purposes (is this possible for every situation?)

Feedback Management A better solution is to turn down only a narrow band, where feedback would have occurred. Better still is to limit the maximum gain in this band, such that an automatic hearing aid is free to vary the gain up and down, but never exceed this preset maximum gain. Dillon (2001)

Feedback Cancelling  External leakage path + - Internal feedback path An even better solution is to cancel the external feedback via an internal feedback path using a filter that has the same gain but opposite phase. Of course, when the jaw moves, the leakage path may change, so the hearing aid needs to continually monitor the input and output, and change the characteristics of the filter accordingly. Dillon (2001)

Implications conceptual complexity - difficult to understand what the aid is doing complexity & adjustability - many different parameters to adjust to set up the aid lack of user adjustability - some nonlinear aids have no volume control - WDRC, in theory, should do away for the need for it test signal - need to chose the right test signal lack of defined standards - no clearly defined standards for measuring nonlinear aids

Ideal vs reality for testing aids Ideal situation: full test-box & programming facility, ability to turn off “extras”, modulated speech-shaped noise as test signal Likely situation for some (eg outreach or other services?): “old” test-box, no programming facility, can’t turn off “extras”, only continuous pure-tone or swept pure-tone available

Summary Signal processing Compression Fits dynamic range of sounds into comfortable range of hearing AGC Types of compression – output-limiting, WDRC Multi-channel processing Implications conceptual, complexity, test-signals

References Dillon, H. (2001) Hearing Aids, Thieme Sandlin, R.E. (2000) Hearing Aid Amplification, Singular Vonlanthen, A. (2000) Hearing Instrument Technonogy, Singular Venema, T. (1998) Compression for Clinicians, Singular Killion, M.C., Staab, W. & Preeves, D. (1990) Classifying automatic signal processors. Hearing Instruments, 41(8), 24-26 Seewald, R. C (2001), A Sound Foundation Through Early Amplification 2000, Phonak AG, ISBN: 3-9522009-0-5 Seewald, R. C. & Gravel, J.C. (2002), A Sound Foundation Through Early Amplification 2001, Phonak AG, ISBN: 3-9522009-1-3 Standards BS EN 61669:2001 Electroacoustics – Equipment for the measurement of real-ear acoustical characteristics of hearing aids BS ISO 12124:2001 Acoustics – Procedures for the measurement of real-ear acoustical characteristics of hearing aids