Digital Signal Processing Reference
In-Depth Information
5.
CONCLUSIONS
In this chapter, for robust automatic speech recognition (ASR) inside a
vehicle (car), a speech enhancement technique based on blind separation of
convolutively mixed signals is applied. This technique is applicable for
under-determined case and hence, is a more practical approach to use in real
applications such as RASR inside a car as compared to other BSS techniques
that work well when the number of sources is equal to the number of sensors.
The signal enhancement capabilities of this technique are verified using a
measure of improvement in speech recognition accuracy. Our preliminary
recognition results of navigation related speech data that was collected in an
SUV show that a significant improvement in speech recognition accuracy -
15 to 35% can be obtained by using our blind convolutive mixture separation
algorithm. Future work warrants testing of the proposed technique using a
larger data set such as the in-vehicle speech data collected by CSLR,
Colorado University (see Chapter 2). Also, the performance of our blind
convolutive mixture separation can be improved if adaptive beamforming
and mixture separation is combined. We are currently working on this.
Future work also warrants using this combined beamforming and blind
source separation based signal enhancement approach to further improve the
speech recognition performance.
