Presentation on theme: "Voiceprint System Development Design, implement, test unique voiceprint biometric system Research Day Presentation, May 3 rd 2013 Rahul Raj (Team Lead),"— Presentation transcript:
Voiceprint System Development Design, implement, test unique voiceprint biometric system Research Day Presentation, May 3 rd 2013 Rahul Raj (Team Lead), Geeta Bothe, Mahesh Sooryambylu, Ravi Ray, Sreeram Vancheeswaran IBM India Customer: Jonathan Leet (DPS 2013) Instructor: Dr. Charles Tappert
Common Passphrase Background: four possible types of passphrases 1. User-specified phrase, like the user's name 2. Specified phrase common to all users “My name is” from phrase “My name is user’s name” 3. Random phrase displayed on the computer screen 4. Random phrase that can vary at the user's discretion Advantages of a Common Passphrase Simplifies the segmentation problem Allows for careful selection of common phrase to optimize variety of phonetic units for their authentication value Facilitates testing for imposters Permits the measurement of true voice authentication biometric performance Avoids potential experimental flaws 2
Software Used: Audacity & Matlab Audacity Open source audio editing software supports sound recording and editing. Supports resampling and stereo to mono conversion Available all platforms: Windows, Linux, Mac Matlab Signal Processing Toolbox provides industry-standard algorithms and apps for analog and digital signal processing Supports visualizing signals in time and frequency domains, FFT computation for spectral analysis, resampling, and other signal processing techniques. 3
System Architecture 4 Collection and management of Speech Samples in repository Preprocessing and spectrogram Generation Mel Filter Banks and MFCC calculation Automatic segmentation “My name is” portion Automatic Segmentation of phonemes using DTW Feature Vector extraction Pace’s Biometric Authentication System will obtain performance results from the feature vectors
Voice Sample Spectrogram using Matlab Input speech Sample (Mono, Samples/sec) 5 Voice stream collected into 1024 frames Samples are read sliding stream by 512 bytes, maintaining overlap Represent samples of a frame One Frame ~ 23ms since Frame size = 44100/1024 Length of one frame = 1000ms/frame size
Voice Sample Spectrogram using Matlab Represent component frequencies of a frame after applying FFT Frequency Vs Time data Voiceprint Systems CS Spring Batch6 Represent the complete spectral data available for processing Spectrogram constructed out of the above values
Mel-Frequency bands space filters appropriately 7 Corresponds to frequency transform performed by the cochlea of human ear. Mel filters are shown below, 13 lower bands are used for processing.
Segmenting “My Name Is” Speech Waveform indicating the voiced and unvoiced segments Energy vs Zero Crossing plotted for same speech sample Non-voiced segments captures high zero crossing rate(red) and low energy(green) values Voiced segments indicate low zero crossing rate and high energy values Voiceprint Systems CS Spring Batch8 Higher frequency components of ‘z’ sound will have higher energy compared to the other phonemes Diagram shows the automatically Marked Spectrum in Matlab Vertical lines demarcate speech beginning and end of ‘z’
Seven sound units of “My name is” 9
Discrete Time Warp (TDW) Algorithm Segments a Sample into Seven Sounds DTW operates on spectrographic data: amp x freq x time To segment a speech sample into the seven sound units, a sample’s time sequence is "warped" non-linearly against a manually sound segmented sample. Voiceprint Systems CS Spring Batch10 Sample warp path represents the cost matrix and the warped path for the two time series represented long the axes If the warp path passes through D(i, j) then the sample Xi is warped to the point Yi. If there is a vertical section in the warp path, a single point in series X is warped to multiple points of series Y. The decision to find the next point in the warp W(i, j) is:
Feature Extraction Features measurements reduce data & characterize speaker The features extracted: Energy mean and variance in each frequency band over the entire utterance (~13*2 = 26 features) Energy mean in each frequency band within each of the 7 phonetic sounds (~13*7 = 91 features) Voice Fundamental Frequency (F0) – not completed Voice Formant Frequencies (F1-F3) – not completed Feature extractor output is a fixed-length vector appropriate as input to Pace University Biometric Authentication System Note: 13 is the number of frequency bands 11
System Performance 12 Feature SetPerformance Features from entire phrase98.05% Features from seven sounds98.95% Performance was measured on 20 sample utterances from each of 30 speakers, manually segmented into the seven sounds. Receiver Operating Characteristic (ROC) curves were obtained to find the Equal Error Rate (EER) and system performance from two feature sets.