by Amyn M. Amlani
Digital signal processing (DSP) algorithms in hearing aids have become increasingly complex since they became commercially available in the mid-1990s. An informal review of several manufacturers' product lines revealed DSP hearing aids with between two and 24 channels of signal processing. Increasing the number of channels offers several theoretical advantages over single-channel signal processing. For example, multichannel compression systems can better accommodate variations in hearing threshold and dynamic range by providing differing amounts of gain across channels (Villchur, 1973).
Multichannel systems with wide dynamic range compression (WDRC) improve speech intelligibility over single-channel signal processing by providing greater audibility at low-input levels (Moore & Glasberg, 1986; Souza & Turner, 1998). Multichannel compression systems also can be designed to be less susceptible to background noise than single-channel compression systems (Moore, 1990; White, 1986). More recently, a benefit of increasing the number of channels is the availability of acoustic feedback technology (Kuk, Ludvigsen, & Kaulberg, 2002; Olsen, Müsch, & Struck, 2001) and noise reduction technology in digital hearing aids (Edwards, 2000; Kuk, Ludvigsen, & Paludan-Muller, 2002).
Despite the theoretical advantages of multichannel compression systems, increasing the number of independent channels does not always result in improved listener performance over single-channel compression (for a review, see Hickson, 1994, and Souza, 2002). This discussion offers a brief overview of how sound is processed through single-channel and multi-channel compression systems, and the potential perceptual reasons for the disparity between listening performance and an increased number of independent channels in a compression system.
Review of Processing Schemes
In a single-channel compression system, the amount of gain and output applied to the incoming signal is controlled by a single set of controllers (i.e., compression threshold, compression ratio, attack/release time; Figure 1 [PDF]). When the incoming signal is loud enough to trigger the compression controllers—either just below the listener's uncomfortable loudness level or at a level that places the input signal into the listener's residual dynamic range—there is a reduction in gain across the hearing aid's entire bandwidth. A major disadvantage of this signal-processing strategy is that it may be detrimental for listeners who exhibit variations in hearing thresholds and dynamic range across frequencies. This occurs because the reduction in overall gain may cause high-frequency sounds to become inaudible, which can reduce speech intelligibility.
Villchur (1973) was the first to advocate that multichannel compression could compensate for variations in hearing threshold and dynamic range by providing differing amounts of gain across channels. Specifically, the incoming signal is split into different frequency channels (Figure 2 on page 12 [PDF]), with each channel's gain and output controlled by its own set of controllers. Depending on the compression architecture, each channel may be controlled by manipulating compression threshold, compression ratio, and attack/release times independently in each channel or across a small subset of channels.
Restoring OHC Function
Outer hair cells (OHCs) play an active role in the cochlea by mechanically amplifying the stimulus in a highly frequency-selective manner and with a near-instantaneous time constant (Dallos, 1992, 1997). Given the cochlea's physiology, it seems logical to design hearing aid signal processing so that it restores damaged OHC function by mimicking the role of normal OHCs: low-threshold, fast-acting compression in many channels.
Unfortunately, multichannel devices with WDRC are unable to restore the mechanical properties of the cochlea. Figure 3 [PDF] shows the loss of the sharp (active) mechanical tuning of the basilar membrane at 3,000 Hz. OHC damage results primarily in the loss of sensitivity to low-level inputs at the characteristic frequency and a substantially broader (passive) tuning curve. The reduction in sensitivity at low-level inputs requires the signal at the input to the impaired cochlea be amplified to be heard.
At moderate- and high-level inputs, the passive OHC system requires less amplification as the input signal increases. Because of the reduced dynamic range, a common strategy is to provide the impaired cochlea with hearing aids having WDRC, which provides differing amounts of gain over a wide range of input levels. Hearing aids with WDRC, however, cannot restore the frequency selectivity of the normal cochlea, but it provides more audibility than other compression strategies at low-level inputs (Moore & Glasberg, 1986; Souza & Turner, 1998).
Figure 4 [PDF] illustrates the nonlinear tuning properties of the normal cochlea at four different intensities as measured from a single fiber in the auditory nerve of a squirrel monkey. Note that for a 25 dB input, OHC response shows a tuning function on the basilar membrane ranging between roughly 2000 and 6000 Hz with a peak response at the characteristic frequency of 4000 Hz. As the intensity level of the input stimulus increases, the tuning function broadens as OHC response becomes more passive. By the time the input stimulus level reaches 85 dB, the response is saturated, and tuning function now occurs over a much broader region. Therefore, increasing gain at moderate- and high-level inputs results in a broader tuning function, even in a normal cochlea.
In the impaired cochlea, however, the broadened tuning function at threshold, whose function is highly variable across listeners with hearing loss, may resemble the tuning function of a normal cochlea at high-level inputs (Florentine, Buus, Scharf, & Zwicker, 1980; Nelson, 1991). Perceptually, the impact of this tuning function affects recognition of vowels (Richie, Kewley-Port, & Coughlin, 2003; Turner & Henn, 1989) and consonants (Dubno & Shaefer, 1995; Turner, Chi, & Flock, 1999). Research also suggests that perceptual performance is highly variable across listeners (Stelmachowicz, Kopun, Mace, Lewis, & Nittrouter, 1995; Turner, Fabry, Barrett, & Horowitz, 1992).
With respect to hearing aid design, multiple narrow channels will produce broader-than-normal excitation in an impaired ear (Trine & Van Tasell, 2002).
Listener Performance
Given the mechanical properties of the cochlea, how many channels are sufficient to improve intelligibility? Most research in this area suggests essentially no improvement in speech intelligibility performance beyond four channels of signal processing (Barfod, 1978; Byrne & Walker, 1982; Walden, Surr, Cord, & Pavlovic, 1999). Only a few studies have shown improvement with more than four channels (Crain & Yund, 1995; Yund & Buckles, 1995a, b).
Recently, Woods, Van Tasell, Rickert, and Trine (2006) attempted to answer this question by quantifying the number of independent signal-processing channels required to maximize the Speech Intelligibility Index (SII; ANSI, 1997) for low- and high-level spectra or to match targets generated by the Cambridge Method for Loudness Equalization (CAMEQ; Moore, 2000). This was performed for a variety of audiometric configurations, ranging from mild to severe. These authors found that one to five channels were adequate to predict speech recognition performance to be within 5% for 90% of the mild and moderate audiograms, and that three to nine channels were needed to achieve the same level of predicted performance for severe audiograms.
With the CAMEQ method, up to four channels of compression provided sufficient flexibility to predict performance in 90% of all audiograms. To better understand the disparity in findings for increasing the number of channels, the perceptual effects of channel summation and temporal/spectral smearing on intelligibility needs to be addressed.
Channel Summation
Clinically, audiologists use prescriptive formulae to estimate the electroacoustic settings of linear and nonlinear hearing aids. This is achieved by adjusting the settings on the hearing aid to match the prescriptive target. Once the output of the hearing aid matches the target, it is often assumed that the listener will receive optimal performance from the device in everyday listening conditions.
Unfortunately, this outcome is not always the case. Multichannel compression devices differ in the number of channels and channel bandwidth. These differences affect the output, which represents the summed output from all contributing channels (Kuk & Ludvigsen, 1999). In other words, as the number of independent channels in a device increases (and as compression ratio increases beyond 3:1 in each channel), the summation effect increases. For example, Kuk and Ludvigsen (2003) compared the output of four different hearing aids with varying numbers of channels and found that the output of a 15-channel device was nearly 10 dB greater than the output for a single-channel device, and nearly 5 dB greater than the output for two- and three-channel hearing aids. Kuk and Ludvigsen (1999, 2003) further indicate that output differences are less variable when complex input signals are used compared to pure tones.
As the sound pressure level at the output increases, there is greater potential for a decrease in speech intelligibility performance due to rollover. To date, channel summation is considered in at least three nonlinear prescriptive formulae (NAL-NL1, Byrne, Dillon, Ching, Katsch, & Keidser, 2001; CAMFIT, Moore, Glasberg, & Stone, 1999; DSLm[i/o], Scollie et al., 2005). These procedures calculate desired real-ear gain for broadband speech-like signals at various input levels. However, small differences in output have been found across the CAMFIT, NAL-NL1, and the previous version of DSL[i/o] (Cornelisse, Seewald, & Jamieson, 1995). These differences have been attributed to channel bandwidth and the manner in which compression is adjusted for a given listener, and the transfer functions associated with converting broadband stimuli into narrow-band target gain values in prescriptive formulae (Leijon, 2004).
It seems likely, therefore, that differences in listener performance between single-channel and multichannel compression systems and among multichannel compression systems differing in the number of channels may be obscured by a rollover effect. Research in this area is needed to corroborate this speculation.
Temporal/Spectral Smearing
Theoretically, compression in one channel of a multichannel system will reduce the gain only in that channel. A potential disadvantage of reducing gain is the decrease in temporal (time-amplitude) difference between phonemes. Kuk (2002) reported that as the number of channels increases in a hearing aid, temporal differences are further decreased, potentially affecting audibility. More specifically, WDRC hearing aids with fast-acting times compress high-level, low-frequency segments (i.e., vowels) and release fast enough to provide gain to low-level, high-frequency segments (i.e., consonants). Relative to the input, the temporal envelope of speech will then be smoothed or distorted (i.e., smeared), and the consonant-vowel amplitude ratio will be increased at the output.
Temporal fluctuations in the speech envelope provide segmental cues for manner of articulation, voicing, vowel identity, and prosody (Rosen, 1992). The reduction in temporal contrasts is most likely to affect listeners having greater than a moderate hearing loss and those with hearing impairment that rely on these perceptual cues to discriminate between sounds (De Gennaro, Braida, & Durlach, 1986; Moore, 1990; Plomp, 1988).
Temporal fluctuations in the speech envelope can also affect spectral (frequency-intensity) contrasts of phoneme identification, and these spectral contrasts are further decreased as the number of channels increases. The negative effects of increasing the number of channels are most pronounced on those sounds that carry pertinent information in the
spectral domain. For instance, a stop consonant (/p, b, t, d, k, g/) is characterized by its release from articulation resulting in a transient noise burst. Given the place of articulation in the vocal tract, the adjacent vowel will have a falling or rising second formant (Borden, Harris, & Raphael, 2003).
For some listeners with hearing loss, the relative amplitude between the transient noise burst and the formant slope is used as a perceptual cue (Hedrick & Younger, 2001). Reducing gain through a fast-acting compression device increases the amplitude of the stop burst, resulting in /t/ being perceived as /p/ (Hedrick & Rice, 2000) and /g/ being perceived as /d/ (Sreevivas, Fourakis, & Davidson, 1997). Studies have also found reducing gain alters the rise-time generated by the transient noise burst in an affricate (/dZ/, /tS/), which affects phoneme identification (Dreschler, 1988; Jenstad & Souza, 2005).
Cognition
Recently, it has been hypothesized that cognitive ability may interact with different signal-processing schemes (Gatehouse, Naylor, & Elberling, 2003; Lunner, 2003), potentially affecting success with amplification. For example, Lunner studied the relationship between working-memory capacity (i.e., high or low) and the ability to identify and report specific effects of a multimemory experimental hearing aid in the field. Memory 1 of the hearing aid was programmed to process the incoming signal differently when speech was present (increase amplification) and when it was absent (decrease amplification). Memory 2 of the hearing aid was programmed to increase amplification regardless of the incoming signal. Using a questionnaire, subjects reported their performance of the two settings in different listening situations. Results revealed that subjects with higher working-memory capacity were better able to identify and report the specific effects of the aid than those subjects with lower working-memory capacity.
Few studies have assessed the relationship between multichannel signal processing and cognition. Recently, Gatehouse and colleagues (2003) tested 50 experienced listeners with five different hearing aid processing systems: single-channel linear; dual-channel linear; dual-channel compression with fast-acting compression (40 msec) in both channels; dual-channel compression with slow-acting compression (640 msec) in both channels; and dual-channel compression with fast-acting compression (40 msec) in the low-frequency channel and slow-acting compression in the high-frequency channel.
Subjects also completed a variety of cognitive tests and outcome measures. Results showed that listeners with greater cognitive ability were able to derive greater benefit in background noise with devices having fast-acting compression, while listeners with poorer cognitive ability performed better with slow-acting compression. Some hearing aid manufacturers now offer defaults for slower release times in their fitting software for older hearing aid candidates (Souza, 2004). The rationale for providing slower release times is that reduced cognitive abilities may limit the successful recognition of the fast components of speech, particularly in the presence of background noise (Cienkowski, 2003). Data from our laboratory suggest that the type of release time best suited for a listener depends more on cognitive function than on age and hearing loss (Amlani, Ahumada, & Miller, unpublished data). Clearly, there is a need for continued research in this area.
Future Needs
Despite its theoretical advantages, compression in many channels has been shown to be neither superior nor inferior to compression in a single channel. In part this finding is due to differences in fitting techniques, compression architecture, and methodology between studies. However, the inability of clinicians to predict those technological features that will enhance a listener's perception of speech may be the most important cause for this paradox.
There is clearly a need to develop clinical measures that can better assess the tuning properties of the cochlea and assess the cognitive effects of listener perception to alterations in natural speech produced by different characteristics of hearing aid signal processing. At the same time, research that evaluates the physical attributes of compression architecture needs to continue.
Only when the bridge between perceptual assessment and hearing aid architecture narrows will we achieve success in adequately fitting different signal processing strategies.
Amyn M. Amlani is an assistant professor in the Department of Speech and Hearing Sciences at the University of North Texas in Denton. His research interests include hearing aid fitting and procedures, auditory perception, cognition and hearing, and marketing trends in the hearing aid industry. Contact him at amlaniam@unt.edu.
This article was adapted from the September 2007 issue of Perspectives, the Internet publication of Division 6, Hearing and Hearing Disorders: Research and Diagnostics. The original article, together with other articles in the issue, constitute content available for self-study that offers ASHA CEUs.for Division 6 affiliates. For more information, see box above.