1) Affective computing aims to expand human emotional intelligence to machines by creating socially intelligent machines that can respond appropriately according to the situation and interlocutor. 2) There are two main approaches to modeling emotions in affective computing: discrete theories that identify basic emotions like Ekman's six emotions, and continuous theories that describe emotions along dimensions of arousal and valence. 3) Empath's goal is to recognize emotions from speech regardless of language, which presents challenges of combining speech processing with emotion recognition from voice cues alone. Empath is developing methods to extract pitch, intensity, and speech rate from voice samples to train models to classify emotions.

1. Afective computing and
emotions in AI
AKAOKA BADSSI Sara
Empath Inc.
12/07/2017

2. Introduction
AI and technology more and more vital in life and society
Almost every domain field relies on it: medical, entertainment..
computers not only actors in technology but also in society
How to make more sociable machines? Many researchers agree that
emotions are part of the answer

3. Emotional intelligence
In social interactions, emotions carry pieces of information: express one's intentions,
interests in the conversation and state of mind
understanding those emotions can improve the interaction between the parties
Isen: emotions can have an infuence on the thinking and reasoning
some tasks are more suitable for some emotions
greeting happiness, comfort compassion
=> exploit and optimize the abilities of the interlocutor by infuencing their emotion
Not only social context: humans need emotions for their survival and adaptation to society
Ex: fear of dangerous situations makes people avoid the danger
understanding the emotions, their origin and consequences
=
emotional intelligence
Tis intelligence allows an individual to make beter decisions for their social
and professional integration.

4. Afective computing
Domain of human-machine interaction
Goal: expand the human emotional intelligence to the machines
overcome the emotional and social gap between human and computers
create socially intelligent machines capable to respond
approprietaly according to the situation and the interlocutor

5. Afective computing:
Discrete approach
Lots of theories based on both discrete and continuous approaches
positive/negative emotions, primary/secondary...
Discrete theory: Paul Ekman
- 6 basic emotions: happiness, anger, fear, neutral, sadness and disgust
- the rest of the emotions can be computed as a combination of those basic
ones
Strong points: universality of the emotion recognition
a basis of a small number of emotions
Weak points: more negative emotions than positive
multiple expressions for one emotion

6. Afective computing:
Continuous approach
Russell's theory:
All of the emotions can be
described with only arousal and
valence
Strong point: only two dimensions,
theoritically one could extract all of
the emotions with this
Weak point: how to measure those
parameters? Which parameters
correspond to arousal? And
intensity?

7. Empath's challenges
Goal: recognize emotions regardless of the language
Strong points:
●
A lot of researches and theories about affective computing, not much practice
●
a lot of studies been done in speech processing, and we can communicate
with machines
Challenges:
●
Affective computing mostly done on facial expression
The universality that Ekman proved is for facial expression only
●
The speech processing that we know is based on words, not emotions. We
know which parts of the spectogram, of the vocal properties take into account
for speech synthesis or recognition, but not emotions
Challenge: Combine both speech and
emotions

8. Empath's approach
Pre-processing the data:
- pitch
- intensity
- speech rate
However, some more or less major obstacles come in the way:
- how to extract those information accuratly and quickly (real-time)
- choice of the model: Random Forest, NN, LSTM...
- still lots of debates about the accuracy of these findings
- individual characteristics (tone, pitch, natural intensity...)
- context and culture
Need of data: 4 emotions, 5 expressions each, 2
genders, 3 types of voices (child, adult, senior), 1 culture:
240000 samples needed
one solution: adding prior information

9. What's next?
“Soon enough it was discovered that it was difficult to find specific voice cues
that could be used as reliable indicators of vocal expressions.Whereas listeners
seem to be accurate in decoding emotions from voice cues, scientists have
been unable to identify a set of cues that reliably discriminate among emotions.”
(Petri Laukka – Vocal Expression of Emotion. Descrete-emotions and Dimensional Accounts – 2004)
Is it a lost cause then?
Not one right answer, but rather a combination of answers,
provide accurate additional information.
Emotions don't carry the entire message and information, they
carry another type of information, different from the one carried in
speech and words.

10. Thank you for your
attention!