Abstract
Reverberant spaces create major problems for hearing impaired listeners. Reverberation
reduces the sound quality and intelligibility of speech for such listeners, especially if they are
hearing aid (HA) users. Moreover, reverberation reduces the effectiveness of many otherwise
useful signal processing methods, such as speech enhancement algorithms, because
reverberation introduces additional virtual sources and background noise that increase the
complexity of the acoustic signals such algorithms must grapple with.
We propose a novel method called SIRCE that “equalizes” (that is, dereverberates) speech
and other audio signals in complex real-world acoustic environments containing multiple
unknown acoustic sources. SIRCE has been designed for compatibility with, and integration
into, NIH’s Open Speech Platform initiative.
SIRCE’s design comprises three critical innovative features: it is room-centric, sensor
image based, and listener aware. Room-centric means that important components of the
system are embedded in the acoustic space itself, rather than residing in the user’s ear (the
hearing aid). There are many advantages to placing sensors (microphones) and processing
components in rooms rather than ears: reduced cost, increased processing power, relaxed form
factor constraints, practical deployment of more than two microphones, and easy sharing of
processing power and computational results between users. A room-centric design makes
particular sense for an equalization system, because reverberation itself is room-specific.
Sensor image based means that SIRCE calculates the acoustic image of each active
source in each sensor. Sensor image extraction (“SIX”) is our innovative contribution to the
active field of blind source separation (“BSS”). Sensor image extraction determines what the
response of each microphone would be to each source in isolation even when multiple sources
are always active simultaneously. SIX computes multiple independent images of each acoustic
source (one for each microphone), whereas typical BSS algorithms only generate a single
estimate for each source. This is important because the most effective dereverberation methods
are multi-channel algorithms that require solo or source-separated inputs from multiple
microphones.
Listener aware means that our system employs the signals from the listener’s HA-internal
microphones, a listener-specific acuity profile, and the listener-specified source of interest
(“target”) to determine whether that target is acoustically audible to the listener; whether other
sources are acoustically audible; and the optimal processing strategy and best sensor image to
present to the listener. (When a target is far from the listener and close to a room microphone,
listening to the HA internal mic response is often not the optimal choice!)
In this Phase I project we propose to validate the sensor images SIRCE computes, quantify
its ability to equalize reverberant speech, and estimate the overall improvement in speech
intelligibility SIRCE delivers.
The SIRCE system will help hearing aid users understand speech better in complex
reverberant spaces.