presented by: karan parikh [email protected] towards the automated social analysis of situated speech...
Post on 19-Dec-2015
216 views
TRANSCRIPT
Presented By:Karan [email protected]
Towards the Automated Social Analysis of Situated Speech Data
Watt, Chaudhary, Bilmes, Kitts
CS546 Intelligent Embedded Systems
• Objective• Difficulties faced in achieving the objectives• Dataset Formation• Privacy Sensitive Speech Processing• Analysis• Application• References
Overview
• An automated approach for studying fine grained details of social interaction and relationship• Local behavior and global structure of the group to be analyzed • Conversation characteristics of a group of 24 people over 6 months are analyzed to study the relationship between conversational dynamics and network position. Objective
Difficulties
• Requires more data than just pure audio recording
• Risk of using audio of uninvolved parties, unethical and illegal
• Risk of people changing behavior if they know that they are being recorded
Dataset Formation
• Data collected from a group of 24 Grad students attending a class.
• All of them carried a MSB(Multi Sensor Board) consisting of tri-axial accelerometer, barometer, microphone, digital compass.
• The MSB was connected to the PDA carried in bag.
• Data was collected over the period of 9 months• Subjects also submitted reports about the
conversations with other subjects for the same month.
Privacy Sensitive Speech processing
• Speech can be modeled in 2 separate components:• Sound generated from the Vocal
Chords• Filter (mouth, nose, tongue)
• Voice can be of 2 types:• Voiced, of fundamental frequency• Unvoiced with no fundamental
frequency.• Intonation, stress and duration are
described by the changes in the fundamental frequency and the energy change during the speech
Privacy Sensitive Speech processing
• Resonant peaks of the frequency response(formants) contain information about the phonemes, which form the basis for words.
• Minimum 3 formants required to reconstruct the words. Once they are removed/not recorded, the word cannot be reconstructed.
• Features that are useful for extracting information from the recorded speech are • Non-initial maximum autocorrelation peak• Number of such peaks• Spectral relative entropy
• The speech is measured in a frames of 60Hz, known as voicing frames• For each frame, the relative entropy is calculated between the
normalized power spectrum of the current voicing frame and a normalized running average of the power spectra of the last 500 voicing frames.
• The accuracy for detection the conversation ranges from 96.1% to 99.2%
Privacy Sensitive Speech processing
• A subject is marked active for the preceding 20 seconds and the following 1 minute
• If the subject is not marked active, then the person is removed from the conversation.
• This feature is heuristic only to prevent the false triggering of conversation recording.
Analysis
• A network is constructed based on the face-face interaction• An edge is created when there is conversation of 20 minutes or more
between 2 subjects.
• Thus a network is formed using multiple edges.• We then examine whether the network thus formed corresponds to
the social relations from the survey and the feedback given by the subject.
• The average aggregate to agreement with the network came to 71.3%.
Analysis
Correlation between speaking style and strength of lies• Assumption: People’s normal behavior is based more on the regular
interaction partners than by rare partners• Hypothesis : people change their way of speaking less when
interacting in their strong ties.• Time spent in conversation by persons I and j estimated to test the
hypothesis.• 4 features that are to be measured of subjects speaking style:
• Rate• Pitch• Turn frequency• Turn Length
Analysis
• bi/j – Mean of i’s speech feature while talking to everyone except j
• bi->j - Mean of i’s speech feature while talking j
• si – Standard Deviation of i’s feature
• dij = |bi/j - bi->j|/si , the amount of feature, that i’s speech changes, when in conversation with j.
• cij = time spent by i and j in conversation.
• Correlation between cij and dij can be measured out.
• The negative correlation : the more the people talk with each other, the less they change their speaking style.
• The table listed below supports the above mentioned hypothesis:
Analysis
Correlation between change in speech features and tie strength
Analysis
Correlation between speaking style and strength of lies• Assumption: People’s normal behavior is based more on the regular
interaction partners than by rare partners• Hypothesis : people change their way of speaking more when they are
speaking with a person who is more central to the network.• The length of the edge for i->j is calculated using (1 – cij) where cij is
that time spent in conversation between I and j. • The centrality of the person is calculated by calculating the
multiplicative inverse of the mean distance of all the points from i.• For no conversation, the edge length will be infinite and for longer
conversation, the length will be shorter.• dik is the change in the feature of all k who speak with i. The higher
the incoming mean, the more the people change their style talking to i.
• Thus dik is correlated with the centrality of the subject.
Analysis
Correlation between change in speech features and tie strength
• More than sociological applications• More than just the identity of the call, the content of the conversation
can be noted and the interruptibility can be predicted.• Can be used in the mobile phones, to divert the incoming calls or
messages to the voicemail during important conversation• Can used to study difference in conversational characteristics of
male/female, people across different geographical regions.
Application
Questions?
• [Donovan, 1996]• http://www.mssociety.org.uk/• Danny Wyatt, Tanzeem Choudhury, Henry Kautz ,” Conversation
Detection and Speaker Segmentation in Privacy-Sensitive situated speech data “
• Jeff Bilmes, Danny Wyatt, Tanzeem Choudhury, Henry Kautz “
References