Two new developments in speech pattern processing hearing aids will be
described. The first development is the use of compound speech patter
n coding. Speech information which is invisible to the lipreader was e
ncoded in terms of three acoustic speech factors; the voice fundamenta
l frequency pattern, coded as a sinusoid, the presence of aperiodic ex
citation, coded as a low-frequency noise, and the wide-band amplitude
envelope, coded by amplitude modulation of the sinusoid and noise sign
als. Each element of the compound stimulus was individually matched in
frequency and intensity to the listener's receptive range. Audio-visu
al speech receptive assessments in five profoundly hearing-impaired li
steners were performed to examine the contributions of adding voiceles
s and amplitude information to the voice fundamental frequency pattern
, and to compare these codings to amplified speech. In both consonant
recognition and connected discourse tracking (CDT), all five subjects
showed an advantage from the addition of amplitude information to the
fundamental frequency pattern. In consonant identification, all five s
ubjects showed further improvements in performance when voiceless spee
ch excitation was additionally encoded together with amplitude informa
tion, but this effect was not found in CDT. The addition of voiceless
information to voice fundamental frequency information did not improve
performance in the absence of amplitude information. Three of the sub
jects performed significantly better in at least one of the compound s
peech pattern conditions than with amplified speech, while the other t
wo performed similarly with amplified speech and the best compound spe
ech pattern condition. The three speech pattern elements encoded here
may represent a near-optimal basis for an acoustic aid to lipreading f
or this group of listeners. The second development is the use of a tra
ined multi-layer-perceptron (MLP) pattern classification algorithm as
the basis for a robust real-time voice fundamental frequency extractor
. This algorithm runs on a low-power digital signal processor which ca
n be incorporated in a wearable hearing aid. Aided lipreading for spee
ch in noise was assessed in the same five profoundly hearing-impaired
listeners to compare the benefits of conventional hearing aids with th
ose of an aid which provided MLP-based fundamental frequency informati
on together with speech+noise amplitude information. The MLP-based pat
tern element aid gave significantly better performance in the receptio
n of consonantal voicing contrasts from speech in pink noise than that
achieved with conventional amplification and consequently, it also ga
ve better overall performance in audio-visual consonant identification
. The signal based on the response to noise of the MLP-based fundament
al frequency extractor was preferred to the noise itself by all five l
isteners. The extension of this pattern classification approach to a w
ider range of acoustic environments and to further acoustic speech fac
tors promises to bring major practical advances in hearing aids for th
e profoundly hearing impaired.