Digital Signal Processing Reference
In-Depth Information
of the signals captured by distributed microphones have been used to ap-
proximate that of the close-talking microphone, through linear regression
[4]. In addition, we have employed MRLS technique for speech recogni-
tion in vehicles and have shown its effectiveness in improving the accuracy
of noisy speech recognition. Through the experiments, we also found that
further improvement of the recognition accuracy can be achieved if the
regression weights are trained for each speaker and/or a particular in-car
sound condition that is mainly set the vehicle itself. These scenarios in-
clude the background music from the audio system, windows are closed
or open, noise from fan/A.C., and the speed of the vehicle. It is worth
noting that the computation of regression weights regression weights for a
given speaker at enrolment is not difficult. Whereas, changing the weights
in order to adapt to the driving conditions is not easy.
The aim of this study is to improve the MRLS so that regression weights
can be changed adaptively to the in-car noise conditions. For this purpose,
we attempt to benefit from distributed microphones for capturing the
spatial distribution of noise sounds.
The rest of the paper is arranged as follows. First, in Section 2, we de-
scribe the in-car speech corpus recorded using distributed microphones.
The basic idea of MRLS and its extension to the adaptive method are
described in Section 3 and Section 4, respectively. In Section 5, experi-
mental evaluations and their results are discussed. Section 6 is a summary
of this paper.
2. MULTIPLE REGRESSION OF LOG
SPECTRA
2.1 Two-dimensional Taylor-expansion of
log-spectrum
Assume that speech signal
at
microphone position is give by
a mixture of the source speech
and the noise
convolved with the
transfer functions to the position,
and
i.e.,
as shown in Figure 19-1. Assume also that the power spectrum of
is
given by the 'power sum' of the filtered speech and noise, i.e.,
