Directional Mics & Digital Noise Reduction Speech in Noise + Hearing Aids = Problems in Noise Why? Ted Venema PhD Recall: Two things we must do for hearing loss: 1. Improve audibility a gain issue achieved with Compression 2. Improve signal to noise ratio achieved with Dmics attempted (but not achieved) with DNR! 1
What it can do to otherwise good people Compression: A Gain Issue Designers of Compression actually look like this OLC for Neural SNHL WDRC for Sensory SNHL Loudness Growth: Normal Hearing vs Sensorineural Hearing Loss Output Output Too Loud 120 10:1 120 Loud 100 100 2:1 Comfortable 80 80 Soft 60 60 Very Soft 0 20 40 60 80 100 Input 0 20 40 60 80 100 Input 0 10 20 30 40 50 60 70 80 90 100 Figure 6-1, Venema T., Compression for Clinicians 3 rd ed Plural Publishing 2017 2
Output Limiting Compression and MPO Adjustment Wide Dynamic Range Compression and Gain Adjustment Output 120 OLC 10:1 Output MPO Adjustment Output 100 WDRC 2:1 Output TK Adjustment 100 80 80 60 60 40 0 20 40 60 80 100 Input Input 0 20 40 60 80 100 Input Input Figure 7 4, Venema T., Compression for Clinicians 3 rd ed Plural Publishing 2017 Figure 7 6, Venema T., Compression for Clinicians 3 rd ed Plural Publishing 2017 Output Limiting vs WDRC: Displayed as Frequency Responses Adjustable Kneepoints Vertically & Horizontally Output Gain 40 db 60 db Gain 40 db 60 db 80 db 80 db Frequency Frequency Input Figure 7 7, Venema T., Compression for Clinicians 3 rd ed Plural Publishing 2017 3
A Multi Kneepoint Input/Output Function ADRO Application of Comfort & Audibility Rules Output Output limiting compression Distribution of Output constantly changes w/listening environment But criteria remain constant WDRC Linear gain Expansion 25 45 65 80 Input 100% Probability of having some output 0% Audibility boundary 30% rule Specified Range of Output db SPL Output Level in one Hz Channel Comfort boundary 90% rule Figure 10 5, Venema T., Compression for Clinicians 3 rd ed Plural Publishing 2017 ADRO: I/O Function Comfort rule applies Unaided Input Speech WDRC 100 Audibility rule applies 90 More Linear Gain Output 80 70 Background noise rule applies Less Linear Gain 60 30 40 50 60 70 80 Input Figure 10 7, Venema T., Compression for Clinicians 3 rd ed Plural Publishing 2017 Aided Output Speech Figure 10 2, Venema T., Compression for Clinicians 3 rd ed Plural Publishing 2017 4
Unaided Input Speech ADRO Normal Input Dynamic Range Aided Output Speech Figure 10 3, Venema T., Compression for Clinicians 3 rd ed Plural Publishing 2017 Figure 10 8, Venema T., Compression for Clinicians 3 rd ed Plural Publishing 2017 Reduced Input Dynamic Range with SNHL WDRC Does This Figure 10 9, Venema T., Compression for Clinicians 3 rd ed Plural Publishing 2017 Figure 10 10, Venema T., Compression for Clinicians 3 rd ed Plural Publishing 2017 5
ADRO Does This Hair cell damage & Speech in Noise Loss of Outer hair cells dulls the traveling wave peak soft sounds no longer naturally amplified Loss of Inner hair cells mixed speech & noise sent on to brain hearing aids make mixed up sound louder Figure 10 11, Venema T., Compression for Clinicians 3 rd ed Plural Publishing 2017 Hearing Aids Make Soft Compromised Sound Into Louder Compromised Sound 6
Solutions for Speech in Noise Presently, there are two... Directional mics objectively improve speech/noise performance Digital noise reduction subjectively enhances comfort in noise Signal to Noise Ratio (SNR) SNR = level of signal (speech you want to hear) versus level of competing noise Eg. 70 db speech in 60 db noise = +10 SNR 60 db speech in 60 db noise = 0 db SNR 60 db speech in 70 db noise = 10 db SNR Typical Face to Face SNRs Pearsons KS, Bennett RL, Fidell S: Speech Levels in Various Noise Environments. EPA Rep 68 01 2466. Environmental Protection Agency, Washington DC, 1977. When noise is 55 db SPL, people speak at about 61 db SPL +6 db SNR: easy When noise is 65 db SPL, speech often 68 db SPL +3 db SNR: harder For those with normal hearing... Speech & noise have to be of similar intensity to understand 50% of the speech +5 db SNR yields 100% speech recognition 5 db SNR is very difficult When noise is 75 db SPL, speech often 74 db SPL 1 db SNR: challenging 7
For mild to moderate SNHL, however... Magic number is around 5 directional mics try to hit this number +5 db SNR gets 50% speech recognition directional mics try to hit this number +10 db SNR yields best speech recognition results in up to 10% speech improvement Basically, an additional 5 db SNR is required for mild moderate SNHL In Summary Each additional 1 db SNR (says Killion) results in up to 10% speech improvement But even if each additional db SNR results in only 5% speech improvement Dmics improve SNR by about 2 3 db Small db Increases Mean Everything According to Dillon, Hearing aids 2012 pg 7: for every 10 db HL, a 1 3 db SNR increase required to maintain intelligibility According to Taylor & Mueller Fitting & Dispensing Hearing Aids, pg 297: every 1 db SNR increases client benefits According to Valente June 29, 2015 https://www.audiology.org/news/cros and bicros hearing aids CROS & BICROS not so good; they: ruin Inter aural loudness difference ruin inter aural time differences Brain has to compare/contrast inputs between right & left ears Better to simply increase SNR 8
Killion; Seminars in Hearing 2002, Figure 3 SNRs for Various Degrees of HL Killion, Hearing Review, Dec 1997 HL (PTA) SNR Required 30 4 40 5 50 6 60 7 70 9 80 12 90 18 Pre Digital Directional Microphones Pre Digital Directional Microphones On BTEs only 1 mic, 2 ports could switch on/off Then came Killion 1998 year after 1 st digital HA ITE Dmic 1 mic, 2 ports cheap but effective omni mics found on ITEs routinely on/off 9
Today s Directional Microphones Directional Microphone Function Routinely have 2 omni mics found on all styles relatively quieter separate program In Any Microphone Sound Moves Diaphragm Sounds Hitting Both Sides Cancel Each Other Out Source Source Source Diaphram Diaphram 10
Directional Mic Function When Sounds Come From Front Directional Microphone Function When Sounds Come From Front Front Direction of incoming sound Rear Front Rear ) ) Time Delay ) ) ) Filter The Diaphragm moves Sound From Behind Sound is cancelled by hitting both sides of diaphragm Directional Microphone Function When Sounds Come From Rear Front Direction of incoming sound Rear Front Rear ( ) ( ( (( ( ( Filter slows sound Acoustic Time Delay Network The Diaphragm cannot move 11
Some Perfectly Rounded Symmetrical Polar Plots: Omnidirectional Cardioid Supercardioid Hypercardioid Directional Microphones: Polar Plots Quantifying polar plots Directional index (DI) ratio of sensitivity to front sounds compared to other surrounding sounds Omnidirectional mics don t favor front sound they have DI of 0 db Hearing aid directional mics try to achieve a DI of about 5 Fig 8-6, Venema, T. Compression for Clinicians 2 nd edition, Cengage 2006 Directional Index (DI): Serves to: Predict the SNR improvement for speech Today we re mainly fitting with RICs or Thin Tube BTEs That might be provided By a specific directional microphone 12
Many of These Have Large Vents Venting reduces low Hz gain where Dmics most effective Open fit BTEs with vented molds reduce gain below 750Hz Open fit BTEs with non occluding molds reduce gain below 1500Hz Dmics & Spacing In dual microphone (twin omni directional) Dmics: Signal subtraction process produces internal mic noise Closer mic spacing further increases internal noise, But also increases high Hz directionality Normal Dmic spacing of about 6 10mm: Gives compromise b/w noise & high Hz directionality Less spacing (5 8mm) increases high Hz directionality From about 4000 to 6500Hz Summary, closer microphone spacing gives: Less Dmic effect, but higher Hz s of directionality Automatic vs Adaptive Directionality Both are used with twin omni directional microphones not statistically better but good for poor manual dexterity and for those who cannot tell when to use what feature Automatic: hearing aid goes from Omni to Dmic by itself depending on listening environment no need for client to manually switch ReSound s Asymmetrical Dmic Fitting Binaural Omni directional: Not very intelligible Binaural Directional: Decrease audibility For side & behind Asymmetric Fitting: Side & behind audible Can shift attention to anyone Adaptive: hearing aid changes various polar plots depending on listening situation also shifts polar plot nulls to origin of noise depending on noise source direction 13
Beamforming: Making Directional Mics Better Yet Mics with more than 2 ports eg. 3 or more mics Killion s ArrayMic TM heart is in right place DI s are about 7 10dB! But would you wear it? Photo provided Courtesy M. Killion Photo provided Courtesy M. Killion Digital Noise Reduction Noise reduction: Implementation fraught with flaws Problem: Speech & noise are mixed together 14
3600 Critical Speech Cues on Spectrogram /ba/ /da/ /ga/ Speech in Quiet Hz 2400 1200 Speech in Noise Frequency 0 0 300 0 30 Time in msec Figure 9 8, Venema T., Compression for Clinicians 3 rd ed Plural Publishing 2017 Figure 9 9, Venema T., Compression for Clinicians 3 rd ed Plural Publishing 2017 Types of Digital Noise Reduction Spectral subtraction Phase cancellation Spectral enhancement Spectral Subtraction 1. Spectrum of Speech & noise together is measured 2. During pauses in conversation, spectrum of noise estimated 3. Spectrum of speech & noise minus spectrum of noise 4. Theoretically, this leaves just spectrum of speech Problem: noise spectrum is wide intersects with speech spectrum So, removing noise removes some of speech! 15
1. Speech Plus Noise Spectrum Speech with its 6db/octave roll off Narrow bands of noise 2. Noise Spectrum b/w Pauses of Speech So, subtract this 75 75 db SPL 50 db SPL 50.1 1. 10 Frequency (khz).1 1. 10 Frequency (khz) 3. From Speech + Noise Spectrum Leaves this Speech Spectrum Really Isn t Too Badly Altered from Original 75 75 db SPL 50 db SPL 50.1 1. 10 Frequency (khz).1 1. 10 Frequency (khz) 16
Problem is, Noise Spectrum is Often Wide Intersects with Wide Speech Spectrum So, this combined Speech + Noise Spectrum Minus the Wide Noise Spectrum Measured during pauses in speech 75 75 db SPL 50 db SPL 50.1 1. 10 Frequency (khz).1 1. 10 Frequency (khz) Ain t gonna leave you with much speech! 75 Phase Cancellation 1. Exact time waveform (not spectrum) of noise is measured 2. Inverted noise phase added to original noise waveform cancels noise db SPL 50 This phase +.1 1. 10 Frequency (khz) Opposite phase = 17
Phase Cancellation used in noise reduction headphones Why not in hearing aids? Because in headphones: speech sent directly to eardrum from headphone noise sampled by microphone outside of headphone Digital hearing aids: do not have this luxury both speech & noise picked up by outside microphone Why Phase Cancellation Can Work in Headphones But Not in Hearing Aids Noise from outside leaks into ear canal and mixes with speech Speech enters directly from headphone Noise from outside picked up by microphone and inverted in phase THIS CANCELS OUT THE NOISE Spectral Enhancement 1. Digital algorithm detects spectral speech cues in noise such as vowel formants or high Hz sibilants Challenge for Spectral Enhancement, however is the high Hz Consonants; Why? In noise: valleys b/w peaks of speech are filled w/noise peaks thus less prominent Low Hz vowels more intense so these still stand out; easier to enhance these 2. Deliberately enhance or amplify these spectral speech cues Just a different approach from noise reduction 18
The high Hz Consonants are what you want most! In noise: valleys b/w peaks of speech are filled w/noise peaks thus less prominent Low Hz vowels more intense so these still stand out; easier to enhance these Today s Digital Hearing Aids Use: Amplitude Modulation (sometimes a combo of Hz & Duration modulation too) DSP algorithms characterize sound waves in each Channel noise waveform has fairly flat waveform over time speech waveform fluctuates rapidly If noise sensed in a channel then gain reduced some 5 20dB Amplitude Modulation Distributions of Intensity Speech versus Background Noise 100 Speech (single talker) Abnormal (non parametric) distribution Mean = Median = Mode / / Background Noise: Less Modulation % time intensity is at some particular db level 50 Noise Typical Bell curve Mean = Median = Mode Speech in Quiet: More Modulation Figure 8 12, Venema T., Compression for Clinicians 3 rd ed Plural Publishing 2017 0 db SPL Figure 8 13, Venema T., Compression for Clinicians 3 rd ed Plural Publishing 2017 19
Abnormal Distribution for Speech Makes Mean Intensity Not in Centre of Long Term Average Speech Spectrum 75 Mean Average Noise Reduction with One Channel Would Reduce Gain Over All Speech Hz s! Frequency Response with no DNR db SPL 50 25 12 db 18 db } Range 100 1000 10,000 Frequency (Hz) Figure 8 14, Venema T., Compression for Clinicians 3 rd ed Plural Publishing 2017 Gain (db) e l Frequency Response with DNR ng un o r e l a J m ng d b i ch sh zv n o r With a 5 15 db DNR gain reduction a p h g sh k all speech sounds drop J m d b i zv p h gch s f k th s f th 125 500 1000 1500 2000 2500 3000 3500 4000 4500 5000 5500 6000 6500 7000 Figure 9 10, Venema T., Compression for Clinicians 3 rd ed Plural Publishing 2017 Noise Reduction with Many Channels Would Reduce Gain Over Smaller Hz Regions Frequency Response with no DNR Speech sounds that dropped Frequency Response with DNR Digital Noise Reduction + Directional Microphones Gain (db) o r u a e l b i ng n p h J m d zv ch sh g k 125 500 1000 1500 2000 2500 3000 3500 4000 4500 5000 5500 6000 6500 7000 Center Hz of Each Band Figure 9 11, Venema T., Compression for Clinicians 3 rd ed Plural Publishing 2017 f s th Noise reduction algorithms give subjective comfort to client Directional microphones gives objective improvement in speech reception Together they make a good team a twin headed approach to speech in noise 20
Ted Venema Hearing Instrument Practitioner Program Douglas College Coquitlam BC Email: venemat@douglascollege.ca 21