Project Description - Robust Speech
Recognition
Robust Speech Recognition by Noise Immunization
Developed a robust speech recognition system using neural network. Using
the multi-layer perceptron (MLP) neural network as a robust classifier
and a modified backpropagation training algorithm, noise immunity is
achieved for SNR levels down to 5 dB while maintaining high recognition
accuracy. The use of noise immunization technique and the correlation of
performance with the order of data presentation for network training are
studied.
A word spotting system is developed to recognize the keyword 'collect'
corrupted by white Gaussian noise in continuous speech.
Robust Recognition for Mobile Communication Applications
Developed robust speech recognition techniques for voice activated dialing
for cellular phone in a car. Noise reduction techniques including linear
and nonlinear spectral subtraction are implemented before modeling the
speech parameters in the homomorphic domain. The FFT derived cepstral
coefficients are liftered and then Mel-scale warped to generate the
feature vector. Radial Basis Function (RBF) neural Network is used as the
final classifier and its real-time performance is evaluated. Performance
evaluation is carried out for both speaker-dependent and
speaker-independent recognition across -5 dB to 25 dB SNR range with
different front-end processing. The system is evaluated using NOISEX-92
and TIDIGITS noise and speech databases.
Speech Processing and Recognition Algorithms
Developed algorithms for pitch detection, endpoint detection,
feature extraction, and several other processes involved in speech
processing and recognition.
For Details See the Relevant Publications