Professional Documents
Culture Documents
1, MARCH 1998 53
Abstract— A portable sound processor has been developed is determined almost entirely by software, new processing
to facilitate research on advanced hearing aids. Because it is schemes can be developed, evaluated, and modified rapidly
based on a digital signal processing integrated circuit (Motorola and relatively easily [1]. In contrast, major changes to the func-
DSP56001), it can readily be programmed to execute novel
algorithms. Furthermore, the parameters of these algorithms tion of an analog signal processor require electronic circuitry
can be adjusted quickly and easily to suit the specific hearing to be designed and constructed, which can be particularly
characteristics of users. In the processor, microphone signals demanding if the device must be small enough to be carried
are digitized to a precision of 12 bits at a sampling rate of or worn by the user.
approximately 12 kHz for input to the DSP device. Subsequently, The ease of programming general-purpose digital processors
processed samples are delivered to the earphone by a novel, fully-
digital class-D driver. This driver provides the advantages of a comes at the cost of high power consumption, and relatively
conventional class-D amplifier (high maximum output, low power large physical size and weight [2], [3]. Although DSP chips are
consumption, low distortion) without some of the disadvantages beginning to appear in hearing aids packaged in conventional
(such as the need for precise analog circuitry). In addition, behind-the-ear (BTE) and in-the-ear (ITE) enclosures, so far
a cochlear implant driver is provided so that the processor these devices are custom-designed and lack the flexibility
is suitable for hearing-impaired people who use an implant
and an acoustic hearing aid together. To reduce the compu- required in research. While they are programmable in the sense
tational demands on the DSP device, and therefore the power that a set of parameters can be adjusted to suit the individual
consumption, a running spectral analysis of incoming signals user, the processing algorithm is to a large extent embedded
is provided by a custom-designed switched-capacitor integrated in the circuitry, so that major changes are not possible.
circuit incorporating 20 bandpass filters. The complete processor Therefore, there is still a need for a sound processor based
is pocket-sized and powered by batteries. An example is described
of its use in providing frequency-shaped amplification for aid on a general-purpose DSP chip. The processor must be no
users with severe hearing impairment. Speech perception tests larger than pocket-sized, so that users can gain experience
confirmed that the processor performed significantly better than with new processing schemes in everyday situations away
the subjects’ own hearing aids, probably because the digital filter from the laboratory. The power consumption must be low
provided a frequency response generally closer to the optimum enough to ensure that neither the size of the battery pack nor
for each user than the simpler analog aids.
the frequency of battery changes will inconvenience the user.
Index Terms— Digital signal processing, hearing aids, speech The ease of programming and the computational performance
perception. must be adequate for the implementation of inventive schemes
that could feasibly be incorporated into digital BTE or ITE
I. INTRODUCTION instruments by a manufacturer within the next few years. This
article describes such a processor. In the following paragraphs,
A variety of devices are now available to assist people with
a hearing impairment, ranging from simple amplifiers
for those with a mild loss, to sophisticated cochlear implant
details of the hardware design are described. Subsequently, an
example is presented of how the device is fitted to hearing-
impaired people, and finally the results of some recent speech
(CI) and auditory brainstem implant systems for those with
perception tests are reported.
a profound or total loss. Although these devices provide
satisfactory performance for many of their users, few, if any,
II. HARDWARE DESIGN
can restore all the characteristics of hearing to normal. Much
of the current research in this field is focused on developing Previously we have described a portable digital sound
advanced sound processing techniques that are designed to processor, called the P-DSP, that was designed specifically
compensate for the perceptual deficiencies that are commonly for cochlear implant research [4]. It has enabled a range
associated with a loss of hearing sensitivity. of experimental processing schemes to be evaluated by CI
In this work, digital signal processing (DSP) is preferred users. Some of these schemes, especially those that atten-
over analog processing for one predominant reason: flexi- uate interfering noise, are potentially beneficial to users of
bility. Because the functional specification of a DSP device other devices, including conventional acoustic hearing aids.
Furthermore, improvements in CI performance have led to a
Manuscript received June 30, 1997; revised December 1, 1997. gradual relaxation of the criteria for patient selection. There
The author is with the Co-operative Research Center for Cochlear Implant, are now people with some usable hearing receiving CI’s, and
Speech, and Hearing Research, Department of Otolaryngology, The University
of Melbourne, East Melbourne 3002 Australia. many of them obtain benefit from wearing a hearing aid and
Publisher Item Identifier S 1063-6528(98)01984-3. a CI simultaneously in opposite ears [5]. These developments
1063–6528/98$10.00 1998 IEEE
54 IEEE TRANSACTIONS ON REHABILITATION ENGINEERING, VOL. 6, NO. 1, MARCH 1998
Fig. 2. Simplified block diagram of the custom-designed spectrum analyzer chip incorporating 20 band-pass filters.
the ADC sampling rate and the selection of the analog input cochlear implants could be driven simultaneously [4]. In that
signal to be digitized. Typically it is programmed to provide situation, each implanted receiver-stimulator obtained data and
an overall sampling rate of 23 438 Hz, with alternate samples a power supply through independent inductive links operating
being obtained from the microphone input. The remaining at a frequency of 2.5 MHz. Thus there were two radio
samples are obtained from each of the other ADC inputs frequency (RF) amplitude-modulated transmitters in the P-
in a cyclical sequence. There are 23 of these inputs in DSP. However, at present there are very few people with
total (amplitudes from 20 filters, dc level from the reference bilateral cochlear implants. To provide an acoustic output
channel, setting of the volume potentiometer, and the battery from the P-DSP/HA, one of the CI transmitters was replaced
voltage). In this way, estimates of the input spectrum are with an earphone driver, while the other was retained. As
acquired at intervals of about 2 ms, which provides sufficient described below, the associated PLD (Xilinx XC3000 series)
temporal detail of changing spectra for most speech-processing was reconfigured to create a novel type of digital-to-analog
applications. converter (DAC) without the need for extensive modifications
The DSP56001 has limited amounts of random-access mem- to the hardware.
ory (RAM) on-chip. In the P-DSP/HA, the memory size is Many recently developed hearing aids produced by com-
extended by provision of 32 768 words of 24-bit static RAM. mercial manufacturers drive the earphone (often called the
This RAM maintains the program code and associated data receiver) with a class-D amplifier [7]. The principle of the
(much of which is specific to the individual user). Because class-D circuit is that an analog waveform is sampled at a
all this information must be retained even when the main regular time interval, and the instantaneous amplitude mod-
power supply is switched off, the external RAM receives ulates the pulse duration (or duty cycle) of a constant-rate
an uninterrupted supply from a miniature lithium cell when pulse train. The modulated pulse train simply controls the
necessary. The external RAM is loaded from an IBM PC- switching of a steady voltage onto the earphone. Because
compatible computer via a simple parallel interface. The earphones generally exhibit a low-pass frequency response
interface also allows the host computer to execute programs with an abrupt slope at a cut-off frequency of typically 4
in the P-DSP/HA, and to examine and modify RAM contents, kHz, they will effectively demodulate the pulse train to recover
thus facilitating software development. Once the program and the original analog waveform. This depends on the pulse rate
data are loaded into the portable unit, it may be disconnected being much higher than any signal frequency, and in many
from the computer and taken away by the user. class-D amplifiers, a pulse rate of about 50–100 kHz is used.
A front-panel switch allows one of up to three stored The advantages of this technique include 1) very high power
programs to be selected by the user (see Fig. 1). Simple visual efficiency, because the earphone is driven through low-loss
indications are provided when appropriate by a light-emitting switches rather than by a linear amplifier and 2) effectively
diode (LED) on the front panel which is controlled directly by linear operation with little distortion up to the output limits
the DSP program. A second LED is provided to warn the user imposed by the power supply. These features are particularly
when the battery is nearly depleted, and is illuminated even valuable in hearing aids, which require long battery lifetime
when the battery is too flat for normal DSP operation. and power amplification with minimal distortion.
Previous class-D amplifiers are mostly analog circuits, but in
D. Output Drivers a fully digital hearing aid, the opportunity exists for converting
In the original version of the P-DSP, provision was made digital sample values directly into the modulated pulse train
for two programmable logic devices (PLD’s) so that two that drives the earphone. In the P-DSP/HA, this conversion
56 IEEE TRANSACTIONS ON REHABILITATION ENGINEERING, VOL. 6, NO. 1, MARCH 1998
Fig. 3. Functional block diagram of the fully digital class-D earphone driver.
provide a 12 MHz clock for the DSP56001. The DSP unit identical earphone and earmold are used for both the threshold
internally divides this clock by 1024 to set the sampling rate measurements and the fitted hearing aid.
to 11 719 Hz for both the microphone and earphone signals. With the P-DSP/HA, thresholds are measured under the
Thus all clocks are synchronized and related by integer ratios, control of a program running on the host computer, to which
resulting in minimum mutual interference. Another feature the portable processor is attached. The host downloads a
of the design is that it does not require any change to the program into the DSP which generates controlled bursts of
format of the data processed by the DSP unit: a conventional sine waves with rise/fall times of 30 ms. The host program
12-bit two’s-complement data representation is used at both specifies the frequencies, levels, and durations of these tone
input and output of the DSP. In this respect the design may bursts in real-time. In practice, an audiologist can operate the
have advantages over techniques based on, e.g., sigma–delta system in essentially the same way as she or he would use
principles, which require parallel data to be converted to and a standard audiometer. Typically, hearing threshold levels are
from serial bit-streams at a rate much higher than the sampling obtained at half-octave frequencies between 125 Hz and 4
rate. kHz, with tone durations of about 500 ms. The levels are
calibrated in standard units of dB HL by converting from level
measurements made with a 2-cm3 coupler [11]. Thus they may
be compared directly with thresholds measured conventionally
III. HEARING-AID FITTING using an audiometer and headphones.
The flexibility of the P-DSP/HA facilitates its use in the The host program computes the required frequency response
development of a wide range of sound processing techniques. of the aid based on the set of measured thresholds. It then
Examples of some types of processing that might benefit subtracts (in dB terms) the frequency responses associated
people with severe hearing impairment include: attenuation with the microphone and earphone of the aid to obtain the
of interfering noise; enhancement of speech signals; improved response that needs to be provided by the processor itself.
automatic control of loudness; transposition of inaudible fre- A simple example of a suitable DSP program is illustrated in
quency bands; and reduction of feedback oscillations. How- flow-chart form in Fig. 5. Samples of the digitized microphone
ever, it is an essential requirement of all hearing aids to provide signal are first processed by a 127-point finite impulse response
amplification at each frequency that is appropriate for the (FIR) filter. They are then multiplied by a value that depends
individual listener. How this basic requirement is attained in on the setting of the front-panel volume control. Finally,
the P-DSP/HA is described below. they are limited to a range of values that is chosen to
A desirable frequency response for a hearing aid is one that prevent the production of uncomfortably loud sounds, and are
optimizes the audibility of information-bearing components conveyed to the earphone driver. The coefficients for the FIR
of speech. The response must account for the characteristics filter are calculated in the host computer by application of
of the aid user’s residual hearing, because hearing threshold the inverse discrete Fourier transform to the derived target
levels, in particular, can vary greatly among individuals. It frequency response. These coefficients and the DSP program
must also account for some important characteristics of the are downloaded by the host into the P-DSP/HA. Subsequently,
speech signal, such as the shape of the long-term average the P-DSP/HA can be disconnected from the host computer
spectrum, which is approximately flat from 100 to 500 Hz, and operated on battery power by the aid user away from
and declines at a rate of about 9 dB/oct from 500 Hz to 4 the laboratory. The power consumption of the P-DSP/HA
kHz [8]. Furthermore, the dynamic range of speech signals is running this relatively simple program is about 450 mW, which
at least 30 dB [8]. The frequency response of the aid should translates into a battery lifetime of about 24 h.
be designed to amplify that range to a range of comfortable
loudness levels at each frequency for the user. Failure to IV. P-DSP/HA PERFORMANCE
achieve these requirements will usually result in less than Recently five experienced users of monaural hearing aids
optimum intelligibility or quality of amplified sounds, and were fitted with the P-DSP/HA using the above procedure.
possibly rejection of the aid by the user. Their performance in understanding speech was assessed with
Various techniques have been proposed for specifying the both the P-DSP/HA and their own hearing aids. In the tests,
frequency response for each aid user. One of the most widely speech material was presented at an average level of 60 dBA
used currently was developed by the National Acoustic Lab- (based on peak levels measured on a sound-level meter with
oratories of Australia, and is known as the NAL-RP formula the fast time constant). The volume control of the P-DSP/HA
[9], [10]. This formula specifies the frequency response using was adjusted to produce the nominal overall gain specified
only the hearing thresholds measured in the individual at each by the NAL-RP formula. For tests using their own aids,
frequency. The thresholds are normally measured with pure each subject was instructed to set the volume control to its
tones using a standard audiometer and supra-aural headphones. usual position for listening to speech at conversational levels.
However, the advent of digital technology has facilitated However, the speech level in the tests was approximately 5
the measurement of thresholds through the user’s hearing dB below typical conversational levels, and was chosen to
aid itself, perhaps resulting in a more accurately specified emphasize the effects of the frequency response of each aid.
frequency response. This is because several potential sources At low speech levels, a nonoptimum frequency response would
of error in computing the required frequency response from be expected to have a greater effect on intelligibility than
the headphone-measured thresholds are eliminated when the at high levels, because more of the speech signal would be
58 IEEE TRANSACTIONS ON REHABILITATION ENGINEERING, VOL. 6, NO. 1, MARCH 1998
Fig. 6. Speech perception results for five subjects using the P-DSP/HA (solid
columns) and their own hearing aids (hatched columns). The ordinate shows
the percent correct of phonemes in monosyllabic words. Error bars indicate
plus one standard deviation of the mean.