Investigate more robust featuresfor Speech Recognition usingDeep Learning
Independent thesis Advanced level (degree of Master (Two Years)), 20 credits / 30 HE creditsStudent thesis
The new electronic devices and their constant progress brought up the chal-lenge of improving the speech recognitions systems. Indeed, people tend touse more and more hands-free devices that are inclined to be used in noisyenvironments. The evolution of Machine Learning techniques has been very ef-ficient for the last decade and speech recognition system using those techniquesappeared. The main challenge of Automatic Speech Recognition systems nowa-days is the improvement of the robustness to noise and reverberations. DeepLearning methods were used to either improve the speech representations ordefining better distributions probabilities. The problem we face is the drop inthe performance of ASR systems when inputs are noisy. The general approachis to define novel speech features that are more robust using Deep Neural Net-works. To do so we got through different implementations as the incorporationof autooencoders in the MFCC block diagram or the deep denoising autoen-coders with different pre-training methods. The final solution is a system thatbuild more robust features from noisy MFCC. Our input is the demonstrationthat a denoising system using q quantized DDAEs defined by the clustering ofthe training data using K-means is more efficient than one denoising systemapplied to the whole data. The performance gained using such a system is of 2to 3% in terms of phone error rate and might be improved using more trainingdata and better tuned NN parameters.
Place, publisher, year, edition, pages
2016. , 58 p.
, TRITA-EE 2016:043, ISSN 1653-5146 ; 043
Electrical Engineering, Electronic Engineering, Information Engineering
IdentifiersURN: urn:nbn:se:kth:diva-183588OAI: oai:DiVA.org:kth-183588DiVA: diva2:912705
Master of Science - Wireless Systems
2016-03-14, Conference Room SIP, Osquldas väg 10 (Floor 3), Stockholm, 10:00 (English)