Claro Fine-scale Noise Canceler Twenty band noise control Introduction The major complaint of hearing impaired people is difficulty listening in background noise (Plomp, 1978; Festen and Plomp, 1990; Needleman and Crandell, 1995; Killion, 1997). In these situations, speech is often masked by the competing noise, speech intelligibility deteriorates, and greater listening effort is required. Consequently, there is a need for hearing instruments that address noise reduction, improve speech intelligibility in noise, and increase ease of listening. Earlier approaches to noise reduction in hearing instruments often included Automatic Signal Processing (ASP) or Adaptive Frequency Response (AFR). These solutions were essentially level dependent with low frequency gain reduction at high input levels. In the presence of steady state low frequency noise, such as car noise (Dillon, 1996) or for low-frequency noise in combination with high frequency speech (van Tasell et al., 1988; Fabry and Walden, 1990; Fabry et al., 1993; Gravel et al., 1999) these systems proved useful.
Claro Fine-scale Noise Canceler
However, their major disadvantage was a lack of selectivity (most were one channel systems) resulting in too much gain reduction and reduced audibility. With the advent of digital technology, it became possible to create more complex noise suppression algorithms which function more selectively and intelligently. These newer systems estimate the presence of noise, and reduce gain accordingly. Current digital noise cancelers function well in the presence of steady state, narrow band background noise, e.g. traffic, or lawn mower noise. They reduce annoyance and improve sound quality, however, they often decrease gain over a relatively broad frequency range. This can result in the loss of important speech cues (Festen et al., 1993; Ludvigsen et al., 1993), and a reduction in overall loudness. There are three reasons for this decreased audibility and speech intelligibility. Firstly, current noise cancelers are limited by the number of frequency bands: having only a few bands precludes gain reduction in narrow frequency regions where noise is present. Secondly, gain
reduction is applied equally to all bands, without taking into account the frequency importance function for speech. Finally, there is no compensation for the loss of overall loudness due to gain reduction. In Claro a more complex noise cancelling algorithm has been implemented capitalising on Claro’s signal processing in 20 bands.
The Claro solutions for hearing in Noise The Fine-scale Noise Canceler is part of a system approach that offers a comprehensive solution for hearing in noise. (Figure1) comprising: • Fine-scale Noise Canceler. • Adaptive digital AudioZoom (not available in all models). • NoiseAdapt DPP - a Digital Perception Processing algorithm specifically for communication in noise. • Automatic activation of the NoiseAdapt program through AutoSelect, when appropriate (speech in the presence of broadband background noise).
Figure 1 Location of the Fine-scale Noise
QuietAdapt
Canceler (FNC) within the Claro system.
FNC
DPP QuietAdapt
FNC Fine-scale Noise Canceler
DPP NoiseAdapt
NoiseAdapt
Auto Select
2
dAZ Adaptive digital AudioZoom
In general: Noise estimation The general principle of noise cancelers is that the presence of noise is estimated and gain is reduced in frequency regions containing noise. The estimation is based on the knowledge that noise fluctuates less than speech; slow fluctuations are associated with noise and fast fluctuations (or modulations) are associated with speech. Estimation occurs within the speech pauses, in each band separately. The larger the number of bands, the finer the estimation of noise and the need for gain reduction. The Fine-scale Noise canceler (FNC ) analyses the incoming signal in 20 separate bands, and estimates the Speech-to-Noise Ratio (SpNR) in each band. Both a slow, long-term average (about 1 second) and a fast, shortterm average (about 10 to 25 ms), are used. The short term average provides an estimate of noise, while the slow averager estimates the speech and the noise (Figure 2). The averaging is continuous while the FNC is active. The analyses result in SpNR estimation
Noise level
Amplitude ➞
Speech level
Time ➞ Figure 2 Temporal waveform of a segment of speech in a noisy background (upper waveform), and in a quiet environment (lower waveform). The red bars in the upper waveform denote the level of the speech (higher level) and the level of the noise (lower than the speech level).
for each of the 20 bands, which serves as criteria for gain reduction in each band. Noise cancelling in a system with few bands results in unnecessary gain reduction in frequency regions not affected by the noise (Figure 3b). The speech signal will also be affected. The larger the number of bands, the better the spectral resolution for noise cancellation. The Claro Fine-scale Noise Canceler works in 20 bands allowing gain reduction only in regions where noise is present (Figure 3c). The speech signal is therefore less affected and the Speech-toNoise Ratio is optimized. FNC: the principles • The amount of gain reduction in each band is dependent on the SpNR per band. • The amount of gain reduction is also dependent on the frequency importance function for speech.
A
Figure 3 B A
The blue shaded area represents the average speech spectrum. The red bar represents a narrowband noise.
B
In a system with few bands (6 in this example) gain reduction will result in loss of speech information as well as a significant loss of loudness.
C C
Since the FNC works in 20 bands, reduction of the same narrowband noise does not result in significant loss of speech information. In addition, loudness impression can be maintained.
3
Percent contribution to speech intelligibility
Gain Reduction based on the SpNR per band Gain reduction occurs when the SpNR in a given band is 0 dB or worse. The degree of gain reduction is directly related to the SpNR in each band. The most gain reduction is at very poor SpNRs of -15 dB, and the least gain reduction at SpNRs of 0 dB. When the SpNR per band, is better than 0 dB no gain reduction takes place, avoiding unnecessary interference with relatively clean signals, and maintaining sound quality. For SpNRs worse than -15 dB a constant, maximum gain reduction is applied. This is to avoid reducing gain too much in extremely noisy situations.
10
8
6
4
2
Band importance (%)
0 0.1
1
10
Frequency [kHz] Figure 5 The Articulation Index (adpated from Pavlovic, 1984) with band specific weighting factors. The mid-to-high frequencies (1000 - 4000 Hz) have a higher band
Band 1 (0-160 Hz):
importance weighting compared to the very low (200 Hz)
SpNR< 0 dB, gain reduction required
and very high (8000 Hz) frequencies. Adding all the values for each band results in 100% intelligibility.
Band 4 (625-780 Hz): SpNR >0, keep gain constant
SpNR
Figure 4 An example of the per band decision making criteria of the FNC. Band 1 (0-160 Hz) has a SpNR less than 0 dB and in this band gain reduction would be applied. Band 4 (625-780 Hz) has a SpNR better than 0 dB, so no gain reduction is needed.
Gain Reduction based on the frequency importance function for speech In addition to reducing gain based on the SpNR a second criteria is applied. Here gain reduction is based on the importance of a given band for speech intelligibility. It is well recognized that not all frequencies contribute equally to speech intelligibility. The Articulation Index (AI) (Pavlovic,1984; 1989) assigns different frequency regions an importance weighting according to their contribution to speech intelligibility (Figure 5). Frequency regions that do not contribute significantly to intelligibility have a low weighting, while those that are most important for speech understanding have a high weighting (e.g. mid-to-high frequency regions).
4
Loudness A common negative "side-effect” of noise canceling is a loss of loudness. In Claro, the fine resolution of a 20 band system works together with the NoiseAdapt Digital Perception Processing algorithm. This ensures that when gain is reduced as a result of narrowband noise interference, the loudness is affected less than in systems with fewer bands. The overall loudness is therefore more likely to remain at an appropriate level.
Gain reduction based on the relative importance of the speech frequencies – the bars represent the amount of gain reduction per band. FNC will reduce more gain in the low
10
Max. Gain reduction
Figure 6
Percent contribution to speech intelligibility
Based on the frequency importance function it is clear that gain reduction in frequency regions with lower weighting will affect speech intelligibility less than gain reduction in regions with higher weighting. In Claro the FNC takes into account these principles and reduces gain accordingly (Figure 6).
8 6 4 2
10 12 8 4
and very high frequencies 0
where less speech information
20
0 0.1
is located. Least gain reduction
1
1
Frequency [kHz]
occurs in the mid-to-high
The combined rule In calculating the required gain reduction, both rules are applied: the SpNR per band and the frequency importance function for each band. Consider, for example, a very poor SpNR of -15 dB across 200-1600 Hz. Given the FNC gain reduction algorithm, 200 Hz has a relatively low importance function, therefore, maximum gain reduction is applied. Less gain reduction occurs in the higher frequency bands because of their increasing importance for speech intelligibility (Figure 7 ). This approach to noise reduction allows reduction of annoying background noise with minimal effect on speech clarity.
frequency region.
Overview of the Fine-scale Noise Canceler: • FNC works in 20 critical bands allowing fine control of noise reduction. • Gain is reduced selectively, and intelligently, only in those of the 20 bands where a poor SpNR is detected. • The amount of gain reduction varies according to the SpNR. The poorer the SpNR the greater the reduction. • Gain reduction is also weighted according to the frequency importance function. Less reduction is applied in frequency bands important for understanding speech. • The NoiseAdapt DPP algorithm configured for noisy environments. • Audibility is preserved due to narrowband noise reduction and the DPP algorithm for noise.
Despite equally poor SpNR, at different frequencies different amounts of gain reduction are applied. More gain reduction occurs at the very low frequency and progressively less gain
Max. Gain reduction (dB)
Summary Figure 7
20
200 Hz 500 Hz
16
1.6 kHz 12
8
reduction towards the mid frequencies.
4
0
1
-15
-10
-5
0
Speech-to-Noise Ratio, SpNR/dB
5
1
The Claro Fine-scale Noise Canceler optimizes the Speech-to-Noise Ratio through a 20 band system. Undesired noisy signals are reduced with minimal effect on "noisefree” speech regions. The FNC algorithm considers the frequency importance function of speech to ensure that important speech information is preserved. The FNC works together with the NoiseAdapt DPP algorithm for superior speech clarity, comfort, and easy listening in noisy environments, and is an important feature of the Claro solutions for hearing in noise.
Claro Fine-scale Noise Canceler
Bibligraphy
Gravel, J. S., Fausel, N., Liskow, C., and Chobot, J. Children's speech
Dillon, H. Compression? Yes, but
recognition in noise using omni-
for low or high frequencies, for low
directional and dual-microphone
or high intensities, and with
hearing aid technology. Ear Hear.
what response times? Ear Hear.
1999; 20:1; 1-11.
1996; 17:4; 287-307. Killion, M. The SIN report: Circuits Fabry, D. A., Leek, M. R., Walden,
haven´t solved the hearing-in-noise
B. E., and Cord, M. Do adaptive
problem. Hearing Jour 1997; 50:10;
frequency response (AFR) hearing
28-32.
aids reduce 'upward spread' of masking? J Rehabil Res Dev 1993;
Ludvigsen, C., Elberling, C., and
30:3; 318-25.
Keidser, G. Evaluation of a noise reduction method – comparison
Fabry, D. A., and Walden, B. E.
between observed scores and scores
Noise-reduction hearing aids. What
predicted from STI. Scand Audiol
is the fate of the ART (adaptive
Suppl 1993; 38:50-5.
response technology)? ASHA 1990; 32:6-7; 48-51.
Needleman, A. R., and Crandell, C. C. Speech recognition in noise by
Festen, J. M., and Plomp, R. Effects
hearing-impaired and noise-masked
of fluctuating noise and interfering
normal-hearing listeners. J Am
speech on the speech-reception
Acad Audiol 1995; 6:6; 414-24.
threshold for impaired and normal hearing. J Acoust Soc Am 1990;
Plomp, R. Auditory handicap of
88:4; 1725-36.
hearing impairment and the limited benefit of hearing aids. J Acoust Soc Am 1978; 63:2; 533-49.
and Plomp, R. The efficacy of a multichannel hearing aid in which
van Tasell, D. J., Larsen, S. Y., and
the gain is controlled by the
Fabry, D. A. Effects of an adaptive
minima in the temporal signal
filter hearing aid on speech reco-
envelope. Scand Audiol Suppl
gnition in noise by hearing-impaired
1993; 38:101-10.
subjects. Ear Hear. 1988; 9:1; 15-21.
www.phonak.com
28152(GB)/0300 • CU/BGS_FNC_GB/0300 • Printed in Switzerland © Phonak AG All rights reserved
Festen, J. M., van Dijkhuizen, J. N.,