Digital Signal Processing Reference
In-Depth Information
present, the higher costs caused by hardware and processing power are
generally accepted because of the benefits to noise robustness. In this
chapter, only a single channel approach was used.
Speaker-microphone distance
.
The microphone is mounted at the A-Pillar
or at the interior mirror. There the Speaker-microphone distance can be
supposed in a range of 20 to 50 cm for the driver. In those distances the
level of the speech signal should be sufficient if the user speaks loud
enough to obtain at least a Signal to Noise Ratio (SNR) of 5, better 10
dB.
Surrounding noise.
Surrounding noises can be eliminated by the built in
automatic noise reduction of the speech recognizer. The noise reduction
adapts automatically to the present background noise. Stationary or
slowly varying noises such as fan, engine or road noises can be handled.
No adaptation is possible for sudden or strong transient noises such as
screen wiper, radio sound or conversational noise.
Conversational noise (babel speech).
Background conversational noise is
difficult to separate from the commands of the driver, because the
conversation has the same characteristics as the commands. Microphones
with directional characteristics or beam forming microphone arrays
possibly attenuate the disturbing speech signal.
4.1 Automatic noise reduction
The Automatic Noise Reduction (ANR) employs the principle of spectral
subtraction and is included in the signal chain of the analyzer after the
Fourier transform (FFT) of the input signal.
Adaptation of the ANR.
For spectral subtraction it is necessary to know
the spectral noise characteristics. Therefore a voice activation detector in the
time domain (TVAD) marks the pause intervals where no speech is present.
The pause decision in the TVAD is based on evaluation of the signal energy
in relation to two adaptive decision thresholds for speech and silence [3].
Spectra in pauses are averaged to estimate the noise spectrum (Figure 12-3).
This is done using a low pass filter [4].
where is the estimated noise spectrum, is the short time
microphone spectrum in pauses, is the adaptation factor, and
l
the spectra
index. Fast, transient changes such as a door clap do not affect the estimation
but slow changes such as the motor sound during car acceleration do.
