55
Digital Noise Reduction: Understanding Lab and Real World Outcomes Ruth Bentler University of Iowa

Analog NR (1980-90s)

  • Upload
    booker

  • View
    42

  • Download
    0

Embed Size (px)

DESCRIPTION

Digital Noise Reduction: Understanding Lab and Real World Outcomes Ruth Bentler University of Iowa. Analog NR (1980-90s). Early spectral approaches Switch ASP ( means low frequency compression) Adaptive filtering Frequency dependant input compression Adaptive compression TM - PowerPoint PPT Presentation

Citation preview

Page 1: Analog NR (1980-90s)

Digital Noise Reduction: Understanding Lab and Real

World OutcomesRuth Bentler

University of Iowa

Page 2: Analog NR (1980-90s)

Analog NR (1980-90s) Early spectral approaches

Switch ASP (means low frequency compression) Adaptive filtering Frequency dependant input compression Adaptive compressionTM

Zeta Noise BlockerTM

Page 3: Analog NR (1980-90s)

Today’s versions Most are modulation-based with some algorithm for

where and how much gain reduction should occur; At least one other (Oticon) first introduced a strategy

called “synchronous morphology” to determine when noise reduction will occur;

Several are now implementing Wiener filters as well Many also use some mic noise reduction, expansion,

wind noise reduction, and even directional mics as part of the strategy they promote.

Page 4: Analog NR (1980-90s)

Today’s talk Focus on DNR Defined here as modulation-based noise

reduction Difficult to “un-involve” the other noise

reduction approaches currently implemented Circuit noise Wind Noise etc

Page 5: Analog NR (1980-90s)

Let’s focus on the impact of Wiener filtering… Norbert Wiener, Missouri-born theoretical

and applied mathematician; developed filter in the early 1940s, published in 1949

VERY interesting fellow….

Page 6: Analog NR (1980-90s)

Let’s focus on the impact of Wiener filtering… The input to the Wiener filter is assumed to be a

signal, s(t), corrupted by additive noise, n(t). The output, x(t), is calculated by means of a filter, g(t), using the following convolution: x(t) = g(t) * (s(t) + n(t))

…where s(t) is the original signal (to be estimated) n(t) is the noise x(t) is the estimated signal (which we hope will equal s(t)) g(t) is the Wiener filter

Page 7: Analog NR (1980-90s)

With DNR shut off, can observe the “onset” of the Wiener filter (~ 3 sec)

Page 8: Analog NR (1980-90s)

Let’s focus on the impact of Sound SmoothingTM… Intended to reduce negative effect of short transient

sounds, such as a door slamming, or cutlery clattering;

Steepness of the envelope slope used to determine if speech or noise (both have crests or peaks)

Very fast time constants; across multiple channels Evidence to support use (Keidser et al, 2007)

Page 9: Analog NR (1980-90s)
Page 10: Analog NR (1980-90s)
Page 11: Analog NR (1980-90s)

How do ‘classification systems’ fit in here? Many high end products have what are

referred to as “classifiers” to categorize the environment for feature activation;

The classification process is likely to impact the onset of many features, esp DNR automatic/adaptive mic schemes Other speech enhancement strategies

Page 12: Analog NR (1980-90s)

Back to modulation-based DNR Modulation count

Important for speech? Typical of noise?

Modulation depth Plomp studies 0-100%

Page 13: Analog NR (1980-90s)

Time waveform of a random noise

Page 14: Analog NR (1980-90s)

Time waveform of a sample speech signal

Page 15: Analog NR (1980-90s)

Modulation spectra

Speech

Noise

Page 16: Analog NR (1980-90s)

Example of algorithm “rule #1”

Page 17: Analog NR (1980-90s)

Example of algorithm “rule #2”

Page 18: Analog NR (1980-90s)

Frequency (Hz)

125 250 500 1000 2000 4000 8000

Diff

eren

ce (d

B,1

/3oc

tave

)

-25

-20

-15

-10

-5

0

5

GN ReSound (CANTA 770-D)

a

Page 19: Analog NR (1980-90s)

Frequency (Hz)

125 250 500 1000 2000 4000 8000

Diff

eren

ce (d

B,1

/3oc

tave

)

-25

-20

-15

-10

-5

0

5

ICRA SpeechRandom NoiseBabble

Starkey (AXENT II AV MM)

b

Page 20: Analog NR (1980-90s)

Siemens (TRIANO 3)

Frequency (Hz)

250 500 1000 2000 4000 8000

Diff

eren

ce (d

B, 1

/3 O

ctav

e)

-12

-10

-8

-6

-4

-2

0

2

SIREN TRAFFICDINING

Page 21: Analog NR (1980-90s)

Unitron (CONVERSA.NT MODA 10A)

Frequency (Hz)

125 250 500 1000 2000 4000 8000

Diff

eren

ce (d

B S

PL

RM

S)

-10

-8

-6

-4

-2

0

2

4

SNR00 SNR05 SNR10SNR15

a

Page 22: Analog NR (1980-90s)

Sonic Innovations (INNOVA)

Frequency (Hz)

125 250 500 1000 2000 4000 8000

Diff

eren

ce (d

B S

PL

RM

S)

-10

-8

-6

-4

-2

0

2

4

SNR00 SNR05 SNR10SNR15

b

Page 23: Analog NR (1980-90s)

Oticon (ADAPTO)

Frequency (Hz)

125 250 500 1000 2000 4000 8000

Diff

eren

ce (d

B,1

/3oc

tave

)

-25

-20

-15

-10

-5

0

5

a

Page 24: Analog NR (1980-90s)

Starkey J13 Axent AV75 dB

--SPEECH,RANDOM, MUSIC--

Frequency(Hz)

125 250 500 1000 2000 4000 8000

DIF

FER

EN

CE

(dB

,1/3

octa

ve)

-25

-20

-15

-10

-5

0

5

GuitarPianoSaxophone with background musicRandom NoisePlain Speech

Page 25: Analog NR (1980-90s)

DNR: What happens in the time domain?

Page 26: Analog NR (1980-90s)

Siemens (Triano)

Page 27: Analog NR (1980-90s)

Starkey (Axent)

Page 28: Analog NR (1980-90s)

Widex (Diva)

Page 29: Analog NR (1980-90s)
Page 30: Analog NR (1980-90s)
Page 31: Analog NR (1980-90s)
Page 32: Analog NR (1980-90s)
Page 33: Analog NR (1980-90s)

Sonic Natura 2 SE BTE DIR 50dB Flat Loss

NOISE REDUCTION: HIGHOmnidirectional

EXPANSION: OFF 85 dB Speech+ Random+Speech

(0:57,1:52,2:51)

Average RMS power: -43.79dB

Page 34: Analog NR (1980-90s)

Average RMS power: -48.04dB

Sonic Natura 2 SE BTE DIR 50dB Flat Loss

NOISE REDUCTION: HIGHOmnidirectional

EXPANSION: OFF85dB Speech+ Random+Speech

(0:57,1:52,2:51)

Page 35: Analog NR (1980-90s)

Reduction: 4.25dB

Average RMS Speech= -43.79dBAverage RMS Noise= -48.04dB

Sonic Natura 2 SE BTE DIR 50dB Flat Loss

NOISE REDUCTION: HIGHOmnidirectional

EXPANSION: OFF 85 dB Speech+ Random+Speech

(0:57,1:52,2:51)

Page 36: Analog NR (1980-90s)

STARKEY Axent AV 50dB Flat Loss

NOISE MGMT:MAX Omnidirectional EXPANSION OFF FEEDBACK OFF 85dB Speech+ Random+ Speech (0:58,1:53,2:51)

Average RMS power: -30.47dB

Page 37: Analog NR (1980-90s)

Average RMS power: -44.03dB

STARKEY Axent AV 50dB Flat Loss

NOISE MGMT:MAX Omnidirectional EXPANSION OFF FEEDBACK OFF 85dB Speech+ Ramdom+ Speech (0:58,1:53,2:51)

Page 38: Analog NR (1980-90s)

STARKEY Axent AV 50dB Flat Loss

NOISE MGMT:MAX Omnidirectional EXPANSION OFF FEEDBACK OFF 85dB Speech+ Random+ Speech (0:58,1:53,2:51)

Reduction: 13.56dB

Average RMS Speech= -30.47dBAverage RMS Noise= -44.03dB

Page 39: Analog NR (1980-90s)

Average RMS power: -31.96dB

Page 40: Analog NR (1980-90s)

Average RMS power: -29.01dB

Page 41: Analog NR (1980-90s)

Reduction(actual increase)=-2.95dB

Average RMS Speech= -31.96dBAverage RMS Noise= -29.01dB

Page 42: Analog NR (1980-90s)

Data?

Page 43: Analog NR (1980-90s)

Data? Walden et al (2000)

Single-blinded, within subject, crossover design 40 HI subjects

Omni versus directional versus directional + NR Self reported:

Speech understanding: NR+D = D = O Sound quality: NR+D = D = O Sound comfort: NR+D > O

Bottom line: Sound comfort evidence

Page 44: Analog NR (1980-90s)

Data? Boymans & Dreschler (2000)

Single-blinded, within subject, crossover design 16 subjects Lab data: NR = No NR Field trials of 4 weeks (APHAB)

All subscales: NR = No NR Three aversiveness questions: NR> No NR

Bottom line: Some reduced aversiveness

Page 45: Analog NR (1980-90s)

Data? Alcantara et al (2003)

Eight experienced HI HA users wore new aid for 3 months

No improvement for SRTs; no decrement for sound quality while listening to four different kinds of background noise, all in lab

Bottom line: No reduction in sound quality

Page 46: Analog NR (1980-90s)

Data? Ricketts & Hornsby (2005)

14 adults, single-blinded, lab data only 2 speech-in-noise conditions

71 dBA speech, +6 SNR 75 dBA speech, +1 SNR

No effect on speech perception Bottom line: Significant preference for DNR

sound quality in lab (forced choice)

Page 47: Analog NR (1980-90s)

Bentler et al (2007) Lab and field study

25 subjects 3-4 weeks field trials with 4 conditions of NR

Fast onset (~4 sec) Medium onset (~8 sec) Slow onset (~16 sec) Noise reduction turned off

Another 3-4 weeks (with “paired comparison”) of three time constants accessed by memory button

Page 48: Analog NR (1980-90s)

Bentler et al (2007) AV (Aversiveness) subscale showed unaided

and NR-off to be significantly different (i.e., unaided and NR-on had similar aversiveness scores)

Diary entries indicate easier listening Bottom line: Less aversiveness and easier

listening relative to DNR-off, both in lab and in field

Page 49: Analog NR (1980-90s)

Examples from diaries: #05

Off: Traffic, TV too loud On: Could hear in conversations with 20 people

#07 Off: Environmental sounds quite loud and did not notice

with other settings On: Seem to have less background noise

#09 Off: Difficult to hear in noise On: Could hear husband in restaurant and understand

almost everything #12

Off: Background and outside noises seemed louder & overpowering

On: Aid seemed to filter out noises almost to the point that

conversation was too low.

Page 50: Analog NR (1980-90s)

What about kids? Current study underway to assess impact of

DNR on novel word learning, speech perception, and sound quality in young children (ages 4-10)

Evidence (in adults) that novel word learning not impaired (Marcoux et al. 2006)

Page 51: Analog NR (1980-90s)

These and other data summarized: Each company has their approach

Often determined by own philosophy Confined by other features (“overhead”)

The outcomes of those different approaches are very different in both the frequency and temporal domains

Does not appear to alter sound quality, speech perception or word learning

Probably makes listening easier Need to verify DNR performance!

Page 52: Analog NR (1980-90s)

Is it functioning as intended?

Page 53: Analog NR (1980-90s)

Is it functioning as intended?

Page 54: Analog NR (1980-90s)

So what’s a clinician to do? Know your product (whose responsibility??) Verify performance

Probe mic measures of gain/output, watch speech, magnitude and frequency distribution of the gain reduction. Same is possible (maybe even necessary) in the test box.

Also can use music passages, babble noise, etc, to observe effect

LISTEN, listen, listen…

Page 55: Analog NR (1980-90s)

Questions?