{ Daniel Wilkey John Graham CS6998 Given speech, was the speaker - PowerPoint PPT Presentation

Detecting Intoxicated Speech { Daniel Wilkey John Graham CS6998

 Given speech, was the speaker intoxicated?  Interspeech 2011 Intoxication Challenge  Application for field sobriety testing, ignition-guards Background

 ALC – Alcohol Language Corpus  162 total participants: 84 male, 78 female  Participants reached a BAC .28 – 1.75  Read 15 minutes of intoxicated speech  Returned 2 weeks later  Read 30 minutes of sober speech The Corpus

 5400 samples in total, 75 per person  Divided into 3 sets:  Development, Training, Test  Development & Training are labeled with 4368 features  Used cross validation to obtain results The Corpus p2

 Shrikanth Narayanan of UCLA  Global speaker normalization  Normalizing by the sober class  Relative improvement of 7.04% overall  Professor Hirchberg  Phonotactic and phonetic cues  Experiment tests un- weighted average recall… why?  We chose f-measure  Includes recall and precision Prior Research

 Remove extraneous features with WEKA  Info-gain ratio algorithm  MFCC features performed well  No F0-based features near the top Experiment Preparation

 Ignore test set  unlabeled  Down-sampling the training set  Achieved 50/50 ratio of alcoholised to non- alcoholised speech Experiment Preparation

 Global Speaker Normalization (Narayanan)  Insignificant negative change  Sober class normalization (Narayanan)  Insignificant negative change  Gender class normalization  Insignificant positive change  Combining global speaker with gender normalization  10.75% relative improvement in f-measure  Poor performance potentially related to some F0 features being filtered out Normalization Attempts

 Tried retesting data with fringe cases omitted  Fringe case BAC between .08% and .16% proposed by Batliner  We tried .02% to .08%  Difference in data set and threshold  Relative decrease of F-measure by 3.25% On the Fringe

Machine Learning Optimizations

 Varied polynomial kernels  Radial basis function (RBF) Optimizing the SVM

 Varying number  Folds  Iterations Optimization Techniques

 Configuration  SVM kernel n=3  10-fold cross validation  Gender normaliation  Sober class normalization Difficult to compare!! Final Results

 Difficult to compare results  Need better corpus  Extend with GMM super-vectors Conclusions / Extensions